Feature Specific Neural Reactivation during Episodic Memory

Transcript

1 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . Episodic Specific Neural Reactivation during Feature - Memory ,* 1 1 ,2 1 2 , Ahmad Michael B. Bone , Bradley R. Buchsbaum , Fahad 1 Rotman Research Institute at Baycrest; Toronto, Ontario, M6A 2E1 ; Canada 2 Department of Psychology; University of Toronto; Toronto, Ontario, ; Canada M5S 1A1 * Correspondence: [email protected] Lead Contact.

2 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . Abstract the original of features many of the component , n experience of the past When recalling a episode may be, to a the mind’s eye. There is strong evidence onstructed in greater or lesser extent, rec that the pattern of neural activity that occurred during is an initial perceptual experience during episodic recall recreated reactivation is (neural reactivation), and that the degree of However, while we know that correlated with the subjective vividness of the memory. reactivation occurs during episodic recall, we have lacked a way of precisely characterizing the — of a reactiv ated memory. Here w e present a contents — in terms of its featural constituents specific informational connectivity (FSIC) novel approach , feature - , that leverages hierarchical representations of image stimuli derived from a deep convolutional neural network to decode n episodic recall neural reactivation in f MRI data col . task lected while participants performed a - W e show that neural reactivation associated with low - level visual features (e.g. edges), high level visual features (e.g. facial features), and semantic features (e.g. “terrier”) occur throughout the dorsal and ventral visual streams and extend into the frontal cortex. Moreover, we show that reactivation of both low - and high - level visual features correlate with the vividness of the with recognition accuracy memory, whereas only reactivation of low level features correlate s - when the lure and target images are semantically similar. In addition to demonstrating the utility , these findings resolve the relative specific reactivation of FSIC - for mapping feature and high contributions of low - level featu res to the vividness of visual memories, clarify the role - to - of the frontal cortex during recollection, and challenge a strict interpretation the posterior - anterior visual hierarchy.

3 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . Introduction quality of experience: some are our conscious memories for past events have the same all Not — on par with the vague and fuzzy, while others are sharp and detailed sometimes nearly “fidelity” of direct perceptual experience. What accounts for this variability in the sharpness and studying mental imagery, episodic memory, and working “resolution” of memories? Researchers memory have over the last several decades or so converged on the idea that memories are constructed from the same neural representations that underlie direct perception (e.g. Ishai et al. Kosslyn 2005; Polyn et al. 2005; Buchsbaum et al. 2012; Johnson 2002; Slotnick, Thompson, & Wing 2015), a process known as & Johnson 2014; Naselaris et al. 2015; Cabeza, Ritchey & Anderson 2010; Rissman & neural reactivation (Danker Researchers have & Wagner 2012). consistently reported that measures of neural reactivation throughout the dorsal and ventral Kuhl, Bainbridge, visual streams reflect the content of episodic memory (Buchsbaum et al. 2012; Laurent e t al. 2014; Cabeza, Ritchey & Wing 2015), & Chun 2012 ; Johnson & Johnson 2014; St - - including low level image properties such as edge orientation and luminosity ( Harrison & Tong, - level semantic properties Albers et al., 2013; Naselaris et al., 2015), as well as high 2009 ; & Haynes, 2011) ( Reddy, . Moreover, the degree of Cichy, Heinzle Tsuchiya & Serre, 2010 ; neural reactivation has been demonstrated to correlate with the memory vividness (Cui et al. 2007; Johnson et al. 2015; St - Laurent, Abdi, & Buchsbaum 2015; Dijkstra, Bosch & van Gerven ne et al., 2019). 2017; Bo The parallels between perception and memory extend beyond the representational overlap within posterior visual regions. As with perception, visual memory is subject to capacity imilar executive processes, such as constraints (Hesslow 2012), necessitating the engagement of s

4 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . & Miller, 2007; Johansson et al. 2012; Wynn et al. selective Buschman attention (Hebb 1968; 2014 & Pearson, Keogh ; Baddeley, 1988 2016; Bone et al., 2019) and working memory ( ; 2015 ). These executive processes serve to enhance and Pearson, Naselaris, Holmes & Kosslyn, relevant image features within posterior visual regions via - maintain neural reactivation of task ; Nobre ( own projections from the frontal cortex - top et al. Mechelli, Price, Friston & Ishai, 2004 et al. 2014; 2004; Higo et al., 2011; ; Dentico D’Esposito, 2012 Dijkstra, Zeidman, & Lee ). 2017 Ondobaka, Gerven, & Friston, Although there is now strong evidence that a network of frontal cortical areas contributes to visual memory, there is currently a debate ove r the nature of the representations within these task regions . According to one account, frontoparietal regions encode abstract - level representations ), Freedman, Riesenhuber, Poggio, & Miller, 2001 category membership ( such as Miller, 2010; 2013 ), and & & rules (Warden Riggall Lee, Kravitz, & Baker, Postle, 2012; stimulus - response mappings ( Rowe, Hughes, Eckstein, & Owen, 2008 ). However, stimulus - have also been discovered within specific responses prefrontal regions ( Miller, Erickson, & Ester, Sprague & Serences, 2015 ; St - Laurent, l, Rissman, & Wagner 2012 ; Desimone, 1996 ; Kuh , ing the - subregions of with some , ) Abdi & Buchsbaum, 2015 frontal cortex encod both task Mante, Sussillo , dimensional state space ( - specific representations in a high - general and stimulus ) to Shenoy, & Newsome, 2013 Raposo, Kaufman, & Churchland, 2014 ; Rigotti et al., 2013; facilitate higher cognitive functions such as attention ( Bichot, Heard, DeGennaro, & Desimone , 2015 ), working memory ( Ester, Sprague Bizley, Jones, decision making ( ) and & Serences, 2015 ). 2016 & Town Whereas evidence for stimulus - specific representations within the frontal cortex , has been growing rapidly over the last decade, there is still little information about the granularity of sensory features repre sented in frontal cortex, as the tools for detecting such

5 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . representations are just beginning to emerge. specific neural representations has advanced significantly over - The detection of feature lutional neural networks (CNN) inspired deep convo - the past few years with the advent of brain Bengio, & Hinton, 2015 ). Early attempts at identifying and localizing neural activity ( LeCun, associated with specific visual features focused on either high - level sematic/categorical features ( et al., 2008; Walther, Caddigan, Fei Fei, & Hung, Kreiman, Poggio, & DiCarl - o, 2005 ; Meyers ) or low - level features Beck , 2009 ; Reddy, Tsuchiya, & Serre, 2010 ; Smith, & Goodale, 2013 Naselaris et al., 2015) such as edges ( Kay, Naselaris, Prenger, & Gallant, 2008; — limiting findings to a sm all slice of the cortical visual hierarchy. In contrast, features extracted from the during activity over nearly the entire visual cortex layers of a deep CNN have been linked to CNN and cortex perception, with a correspondence between the hierarchical structures of the Eickenberg, Gramfort, ( Yamins et al., 201 4 ; Güçlü and van Gerven, 2015; Wen et al. , 2017; Varoquaux, & Thirion, 2017 ; Seeliger et al., 2018 ) . feature - specific neural and Kamitani ( 2017 ) used this approach to reveal Horikawa reactivation throughout the ventral visual stream during mental imagery. However, because Horikawa were predicting category Kamitani average features, as opposed to image - specific and - features, the study was insensitive to the reac level features due to the large - tivation of lower category variability of lower - level features relative to higher - level features. Moreover, the within - authors’ decoding approach did not account for the inherent correlations between the feature - cted from the CNN. The architecture of feedforward CNNs is designed such that levels extra features from higher layers of the network are composed of features from lower layers, resulting - layer - ese inter layer correlations. Thus, any method that does not control for th in strong inter correlations will be prone to false positives, i.e. falsely detecting reactivation of features from

6 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . (nearly) all levels of the visual hierarchy when only a small subset of the feature levels are - ) and ( et al. (2018) van Gerven d an Güçlü present within a given brain region. Seeliger 2015 developed a method to address this issue that first assigns the layer that best predicts a given voxel/source’s activity to that voxel/source, and then uses the proportion of voxel/sources levels represented within that cortical - er within an ROI to infer the feature each lay assigned to levels that are weakly represented within - region. This approach, however, may overlook feature presented per a given region, due to the simplifying assumption that only one feature level is re voxel/source, resulting in false negatives. specific - To overcome some of these previous imitations in identifying feature - specific informational connectivity reactivation during memory recall, we introduce feature - incorporates a voxel (FSIC), a novel measure that wise modeling and decoding approach & Coutanche (Naselaris et al., 2015), coupled with a variant of informational connectivity ( & Coutanche, 2018 Thompson - - Schill , 2013 ; Anzellotti ). Unlike previous measures of feature specific neural reactivation, our method takes advantage of trial - by - trial variability in the retrieval of episodic memories by measuring the synchronized shifts in reactivation across accou for cortical regions. We demonstrate that this approach eliminates false positives by nting - inter layer correlations while retaining sensitivity to more weakly represented features. specific reactivation across the neocortex during a task - We used FSIC to examine feature requiring subjects to recall and visualize complex naturalistic images. The experiment consisted ls, and three sets of of three video viewing runs (Fig. 1b), used to train the encoding mode alternating encoding and retrieval runs (Fig. 1a). During the encoding runs participants Back task. In the following - memorized a sequence of color images while performing a 1 mory of the images was assessed. and recognition me recall retrieval runs, the participants’

7 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . Feature specific neural reactivation was measured while participants visualized a cued image - then . The participants memory vividness rating grey rectangle, followed by a - within a light ical image during encoding, followed by a rating of their judged whether they had seen the ident confidence in this response. Given the purported role of the frontal cortex in coordinating visual representations within posterior sensory regions ( et Nobre ; 2004 Higo Mechelli, Price, Friston & Ishai, al. 2004; et al., 2011; et al. 2014; ; Dentico Dijkstra, Zeidman, Ondobaka, Gerven, & Lee D’Esposito, 2012 & Friston, 2017 ), we hypothesized that neural reactivation associated with all visual feature - these cortical regions. Beyond — and be synchronized between levels should occur within — also specific visual representations, we were - establishing the cortical distribution of feature connection to memory performance. To this end, we investigated the interested in their and relationship between feature - specific reactivation during recall both subjective (vividness ures. We hypothesized ratings) and objective (recognition accuracy) behavioral memory meas but that reactivation of all feature levels would correlate with the vividness of the recalled image, these features would have the strongest correlation because lower level representations that are (Hebb, arp and intense phenomenology of vivid memories most clearly associated with the sh ). For the relation between reactivation and recognition Kosslyn, Ganis, & Thompson, 2001 1968; level features should not assist in differentiating the encoded - memory, we hypothesized that high d the lure due to the close semantic overlap between the two images (see image an , we hypothesized that only the thus ; Supplementary Figure 5 for example image pairs) level features during reactivation of lower should correlate with recognition accuracy. recall -

8 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . Fi and Visual Features . a) Alternating image encoding and retrieval tasks. gure 1. Procedure back task while viewing a sequence of color - During encoding, participants performed a 1 1) were pants During retrieval, partici photographs accompanied by matching auditory labels. presented label, 2) retrieved and maintained a mental image of the cued with a visually - associated photograph over a 6 second delay, 3) indicated the vividness of their mental image and 5) entered their using 1 e cued item, 4 scale, 4) decided whether a probe image matched th - . confidence rating with respect to the old/new judgement b) Example stills from the two videos shown before the encoding and retrieval tasks. Data from the videos, which comprised a series of eatures were short clips, was used for trai ning the encoding models. c) F or each image , f extracted from layer node activations using the VGG16 deep neural net (DNN). Activations from nd th th convolutional layers, and the last fully connected layer were used, , 7 the 2 and 13 middle - high visual, - corresponding visual, - to low visual and semantic features, respectively.

9 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . Results Behavioral Imagery Vividness Ratings The mean vividness rating over trials, averaged across participants, was 3.04 (SD = 0.35). On vividness = 1, 19.7% as vividness = 2, 48.0% as vividness = average, 2.9% of trials were rated as Participants failed to respond within the three second vividness 3 and 29.4% as vividness = 4. rating period on 0.9% (SD = 2.2%) of the trials. These trials were excluded from all analyses. Old/New Task Accuracy and Confidence Ratings The means for old/new task accuracy and confidence ratings, averaged across participants, were 81.0% (SD = 11.0%; chance = 50%) and 3.46 (SD = 0.30), respectively. On average, 3.1% of ence = 1, 10.9% as confidence = 2, 22.1% as confidence = 3 and 63.9% trials were rated as confid as confidence = 4. The association between accuracy and confidence ratings was significant (β = model with subject and .89, p < .001) (measured using a generalized linear mixed - effects (LME) Participants failed to respond within the three second old/new image as crossed random effects). response period on 1.0% (SD = 1.5%) of the trials, and the two second confidence rating period on 1.8% (SD = 2.3%) of the trials. The former trials were cl assified as incorrect, while the latter were excluded from analyses that incorporated confidence ratings. Neural Reactivation During Episodic Memory Recall Measuring Neural Reactivation Using an Encoding Decoding Approach - - , an encoding memory recall ing To measure neural reactivation dur decoding approach was used

10 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . (Naselaris et al., 2015). In short, encoding models were used to predict the expected neural he correlation a seen or imagined image. T comprising activity in response to a set of features activity measured during between del predictions and was then used to visual mo the recall decode the cued image. Brain activity measured during the encoding runs and the first two video runs were used wise encoding models for e surface cortical to train ach of four visual feature levels: based vertex - - level visual features and semantic level visual features, mid - level visual features, high low - - features. Given recent work showing a correspondence between visual features derived from a n 2015 ; image recognition CNN and the feat ures underlying human vision ( Güçlü & van Gerven, ), the encoding models used features extracted from layer activations Horikawa & Kamitani, 2017 ) to predict neural activity (Fig. 1c). Based on & Zisserman, 2014 in a DNN (VGG16; Simonyan Güçlü & van Gerven the findings (2015), outputs from the units in the second, seventh, of - visual, mid - thirteenth, and sixteenth layers of VGG16 were used as approximations of low visual and semantic cortical features, respectively. visual, high - were accurately characterized by the vertex - wise feature - brain that To identif y regions specific encoding model s, neural activity predicted by the encoding models for each trial and feature Destrieux, Fischl, D ale, & - level were grouped into 148 bilateral cortical Freesurfer ROIs ( ). For each ROI and trial, predictions of neural activity for all encoded images Halgren, 2010 were generated and correlated with the observed neural activity. The predictions were then sorted cued image actual n associated with the by correlation coefficient, and the rank of the predictio was recorded. To make the rank measure more interpretable, the rank was subtracted from the mean rank so that a value significantly greater than 0 indicates neural reactivation (i.e. the cued image could be decoded f recall ). rom neural activity during

11 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . Reactivation for each bilateral ROI . Episodic Recall Neural Reactivation During . 2 Figure column = feature ) was significantly greater than chance and feature level ( Reactivation . throughout the dorsal and ventral visua l streams and within the lateral and orbital frontal cortex The t - during recall . values are thresholded at p < 0.05, FDR corrected. Figure 2 depicts neural reactivation for all cortical ROIs during . episodic recall - Consistent with previous findings (Buchsbaum et al. 2012; Johnson and Johnson 2014; St 2017 ), the ab ility to Laurent et al. 2014; Cabeza, Ritchey and Wing 2015; Horikawa & Kamitani, for all throughout the dorsal and ventral visual streams greatest decode recalled memories was feature levels, with the peak ROIs (Figure 4a; see ROI/Seed Selection for details) located within brain regions of the cortical visual processing hierar chy associated with each feature level. Significant decoding accuracy was also seen in the lateral prefrontal cortex, particularly within inferior frontal sulcus. Overall, our findings indicate widespread neural reactivation the - episodic recall. associated with all feature levels during - Specific Informational Connectivity Feature - re levels throughout the Despite strong findings indicating reinstatement of all CNN featu cerebral cortex, correlations between features from different network layers (Supplementary Figure 1) makes it difficult to independently assess the contribution of each feature level to gher level features are composed of lower level memory reactivation. For example, because hi features, it may be the case that only neural activity associated with low - level visual features is level visual - and low - reactivated in a given ROI, but, due to the correlation between high

12 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . level visual features also appears to be reactivated. This features, activit - y associated with high could explain why reactivation of semantic features was found within the calcarine sulcus vel visual le - which is surprising given the area’s assumed role in low (Figure 2, last column), ). Thus, to assess the independent contribution of each ( processing Güçlü and van Gerven, 2015 feature level to reactivation, it is necessary to statistically account for neural activity associated specific informational - veloped feature with all non - target features. To that end, we de a variant of informational connectivity ( & Thompson — - Schill , connectivity (FSIC) Coutanche - ). FSIC measures the correlation of feature 2013 specific neural reactivation between a seed ROI - levels, thereby enabling the and a target ROI, while covaryi ng out all non - target feature detection neural reactivation associated with a specific set of features. Specific Informational - Simulated Results for Decoding Accuracy and Feature . 3 Figure fMRI data was simulate d (200 simulated subjects; see Methods section) and then Connectivity . run through the processing pipeline for FSIC (see Methods section) to validate the approach. - ROIs only contain features from the indicated feature level. a) Image c lassification performance Correlations between feature - - (rank mea sure) for all combinations of ROI and feature level. levels result in the classification accuracy measure falsely indicating the presence of features that - are not present within the target ROI. b) FSIC results for all combinati ons of ROI and feature - trial memory accuracy across feature level assuming identical trial - levels. A separate seed by - level (the results are also depicted was used for each feature - level corresponding to that feature - Significant agonal). b along the di 2 in Supplementary Figure FSIC results only indicate the level contained within each ROI, except for relatively weak evidence for - presence of the feature

13 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . the presence of adjacent feature levels (e.g. a significant effect associated with mid - level features - Error bars are 90% CIs; * indicates p < 0.05, FDR level ROI). - ound within the low was f . corrected Before applying FSIC to experimental data we first validated the approach with a neural reactivation associated with a specific can detect FSIC to determine whether simulation level, while eliminating fMRI data was simulated for visual feature false positives To this end, - . from the CNN in response to the experimental /outputs activations node using the 200 subjects (see fMRI for details ). Figure 3 a depicts the classification accuracy stimuli Data Simulation results for this simulated data. Despite each ROI representing features from only one feature - , significant effects are present for all feature ssification levels within each ROI. If cla - level accuracy were to be used to infer the representation of features within a given ROI, it would lead all feature - levels. Figure 3b s each to the false conclusion that contain represent ations of ROI - by - trial reactivation depicts neural reactivation results using FSIC as suming identical trial naïve fidelity (i.e. proportion of forgotten features) across feature - levels . In contrast to the classification accuracy method, FSIC accurately identifies neural reactivation associated with of signal smearing to with a small amount only the features p resent within each ROI, albeit trial reactivation fidelity - by - No signal smearing was found when trial . features in adjacent layers levels Figure was assumed to be independent across feature - (see diagonal of Supplementary compare diagonal of figure 4b ( - an assumption that more accurately modeled the off — 2c) so figure 3b likely overestimates the amount of signal Supplementary Figures 2b and 2 — c) s imilar, yet smearing one can expect when applying the technique to real data. Moreover, generally weaker, results were found when the seed ROIs contained an equal proportion of voxels representing each feature - level (seeds in the above results only contained the target , suggesting that - specificity of FSIC is not dependent on the selection of - feature the feature level) ( ). a Figure Supplementary FSIC may level - that only contain the target feature seed ROIs 2

14 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . neural reactivation associated with therefore b greatly improve our ability to isolate e used to when compared to the naïve approach. specific features - uring 4 Figure Specific Informational Connectivity . Seed ROI Weights and Feature d All ROIs are bi a) Seed ROI weights for each feature level. Seed Episodic Recall . ROI lateral. weights are proportional to the decoding accuracy for the target feature level relative to the other during encoding/perception (i.e. the relative accuracy peaks) feature levels . b) FSIC results for neural level. For FSIC, level and target ROI feature - all combinations of seed ROI/feature - seed ROI reactivation (during memory recall ) of each feature level within the corresponding four ; seed ROIs colored feature levels (rows dark blue) was correlated with reactivation of all , within all ROIs except feature levels (columns) target - of the non reactivation controlling for the , The t - values associated with those correlations are indicated with shades of red and . for the seed thresholded at p < 0.05, FDR corrected.

15 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . Figure 4b depicts the results from applying FSIC to fMRI data measured during obtained Specifically, the figure displays the correlation Figure 1a). i.e. the task depicted in ( recall visual n ROIs of each feature level within the corresponding seed ROI s (rows; neural reactivatio for target four all and marked with blue) Figure 4a from feature levels within all other ROIs feature levels. diagonal results indicate the controlling for , all (columns) non - Off - target co diagonal results indicate the correlation - levels, whereas on rrelation between different feature - level (Figure 3 depicts a simulation of the latter). According to our within the same feature - - simulation results, the generally weak correlations within the of f - diagonal indicate that trial - by - levels is independent across feature trial variation in memory reactivation largely c), i.e. level is only from the degree of reactivation - - b 2 (Supplementary Figures one feature a different feature - level. weakly related to the amount of feature react ivation from In contrast, visual, s trong correlations along the diagonal indicate widespread neural reactivation for low - order - visual, and semantic features that extends beyond the occipital cortex into higher - high - dorsal and ventral visual streams as well as the frontal cortex. Reactivation of mid regions of the and level features was, however, primarily limited to the occipital cortex; this difference is not Supplementary Figure . due to ) the relativity small size of the mid - level seed ROI 3 (see Although ; Freedman, Riesenhuber, Poggio, & Miller, 2001 order features ( - expected for higher Carota, - Blagoev, Clark, & Poldrack, 2001 ; Huth, Nishimoto, Vu, & Gallant, 20 Wagner, Paré ; 12 Kriegeskorte, Nili, & Pulvermüller, 2017 ), the widespread presence of low - level visual features order regions (see Table 1) appears to challenge a strict interpretation of the - within higher for mid similar to - level what we observed cortical visual hierarchy, which would predict results visual features.

16 bioRxiv preprint first posted online Apr. 30, 2019; The copyright holder for this preprint doi: http://dx.doi.org/10.1101/622837 . (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a . CC-BY-NC 4.0 International license SE Lower P t value Upper Frontal β ROI (FDR corrected) bound bound 0.117 0.004 ** Middle frontal sulcus 0.021 3.9 0 0.05 0 0.083 0.117 0.004 ** 0 3.9 0.021 0.083 Superior precentral sulcus 0.047 3.89 ** 0.047 0.004 0.021 0.083 Superior circular sulcus 0.115 0.118 0.004 ** Inferior precentral sulcus 0.083 0.021 3.87 0.046 0.114 0.004 ** Superior frontal gyrus 0.077 0.022 3.47 0.039 0.104 ** 0.034 0.004 3.16 0.022 0.068 Anterior midcingulate Superior frontal sulcus 0.008 ** 0.067 0.023 2.97 0.027 0.107 0.093 0.01 0 * Middle frontal gyrus 0.057 0.022 2.61 0.022 0.018 0.023 * Anterior cingulate 0.055 0.022 2.55 0.091 0.022 * 0.09 2.48 0.021 0.053 Short Insular gyri 0.016 0 0.089 * Precentral gyrus 0.016 0.018 2.44 0.022 0.053 0.013 0.083 0.02 0 * Inferior frontal gyrus - 0.048 0.022 2.25 o percular within Table 1 . Low - Level Feature - Specific Informational Connectivity During Imagery the Frontal Cortex . The table lists the significant FSIC results (and associated statistics) within the frontal cortex depicted in the first row and first column of figure 4b. al Reactivation and Neur Behavioral Performance Between Neural Reactivation and Vividness Ratings Correlation specific neural reactivation established, we then assessed - With the cortical distribution of feature specific reactivation relates how feature - To to behav test ioral measures of memory performance. - level result of the reactivation of lower is largely the vividness memory the hypothesis that and mid - Kosslyn, Ganis, & Thompson, 2001 visual features (Hebb, 1968; ), measures of low - l evel features), and high - level and semantic reactivation (higher level reactivation (lower level - - features) were combined (averaged) together, along with the associated ROIs (Figure 5a), forming four separate reactivation measures: one for each unique combination of feature - level - subject correlations between these reactivation measures and vividness was an d ROI. The within then assessed using a LME model, with the vividness rating as the dependent variable (DV), the four reactivation measures as independent variables (IV) bject and image as crossed and the su ,

17 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . random effects - intercept only, due to model complexity limitations) , thereby controlling (random levels and ROIs. Figure 5b depicts the partial regression - for the correlations between feature reactivation measures. Contrary to our hypothesis, coefficients associated with the four lower level features, within corresponding ROIs, appear to - reactivation of - level and higher equally to subjective vividness. approximately contribute level - higher of In addition to the positive partial correlations, we found that reactivation level ROI negatively correlated with vividness. According to the - features within the lower predictive coding account of perception - down connections from neurons that encode high - , top drive neural activity - level features to generate a lower representing level/semantic features model of the expected stimulus , which is compared against the perceptual input to generate an 2005, 20 10; Bastos et al., 2012) . From this Friston, ; & Ballard, 1999 ao R ( error signal represent the perspective, the presence of higher - level features within the lower - level ROI may - top - down inference of lower level features. When reactivation of the perceived low - leve l statistically controlled — as in the above analysis — the reactivation of higher - level features is for incorrect only the is constrained to represent level ROI - features within the lower inferences, i.e. in the encoded image. These incorrect - level features that were not present predictions of low - inferences would result in mental imagery of a generic image associated with the recalled high even if the rate as vivid to which participants were instructed not level/semantic features — many visual detail s . Therefore, the observed negative partial generic mental image contained - - level features within the lower correlation between vividness and neural reactivation of higher . is consistent with level ROI predictive coding account of perception and memory recall a

18 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . Specific Neural Reactivation, Vividness and - Correlations Between Feature . 5 e Figur and Recognition Accuracy . and mid a) ROI weights combining the low - level and high - - relation measuring the subject partial regression coefficients - level ROIs. b) Within - semantic neural reactivation during recall - level and between and vividness for all combinations of feature measuring the relation neural between subject partial regression coefficients - c) Between ROI. for all combinations of feature - reactivation and recognition accuracy (during the Old/N ew task) The error bars are 95% CI; * indicates p(β = 0) < .05, FDR corrected. level and ROI. Correlation Between Neural Reactivation and Recognition Accuracy recognition accuracy during the would selectively correlate e W hypothe sized old/new task that visual due during features episodic memory recall, level - er with reactivation associated with low semantically similar to the original image but differing with respect to its to the lure image being . To low - level visual features test this claim, the same analytical approach described above for the was used, replacing vividness with accuracy as between reactivation and vividness correlation ). (Figure 5c) . Next, we = 0 incorrect = 1, significant correlations were found (correct the DV No - subjects using a similar model to the one used for the within - between the correlation examined subject analysis (except averages, subject and subject and image - the DV and IVs were within random effects). Consistent with our hypoth esis, we found a significant partial were no t used as - correlation between recognition memory accuracy and lower level reactivation within the subject coefficients were pooled together within and between - - corresponding ROI Figure 5d; ( . ) ns when controlling for multiple compariso using FDR

19 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . use neural Th - subject result suggests that only some subjects may successfully e between null The reactivation within early visual regions to enhance recognition memory accuracy. therefore individual difference. To explore be finding subject - within a consequence of this might this possibility, neural reactivation was calculated the correlation between memory accuracy and subject linear model, except subject and image - for each subject individually (using the within were not used as random effec for each ts). The resulting partial correlation coefficients level - correlated with the subjects’ memory separately were then combination of ROI and feature 4 a - c, respectively accuracy . for all trials, lure trials, and ‘old’ trials ( S upplementary F igure s ) lure trials, level features on - and higher - for lower ignificant positive correlations were found S ing that the hypothesized positive within - subject correlation between memory accuracy suggest and neural reactivation may only be evident for su bjects with relatively high recognition y - subject analysis accurac . This possibility was tested using the same model as the original within memory with the highest average seven) - on (half of the twenty on the thirteen subjects accuracy the lure trials d 4 igure F lementary upp S ( ; the thirteen subjects with the lowest average memory performance group, lower - accuracy were also tested : Supplementary Figure 4 e ) . For this high - level features within the corresponding ROI positively correlated with memory accuracy [β = - 1 .08 , p = .0 14 ] . Thus, there is a relationship between low level feature reactivation and recognition memory performance, but it is limited to the higher performing subset of participants. Discussion - the specific feature The primary goal of the current study was to identify level composition of neural reactivation patterns measured throughout the neocortical mantle during a task requiring

20 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . - level, vivid . We found that neural reactivation of low recall of a diverse set of naturalistic images level, and semantic features occurs throughout the cortex, including much of the - level, high - mid A classic theory relating neural reactivation dorsal and ventral visual streams and frontal cortex. mental imagery is to the subjective qualities mental imagery postulates that the vividness of based primarily level visual features, e.g. edges and simple - lower of upon neural reactivation level features and higher - - shapes (Hebb, 1968). We found, however, that reactivation of lower hypothesized that further equally to subjective vividness. We approximately contribute and visually similar images during a visual recognition memory distinguishing two sema ntically - task would benefit from neural reactivation of lower level visual features, particularly within level - early visual regions. Consistent with our hypothesis, subjects with greater lower had greater recognition accuracy. recall reactivation within the early visual cortex during - subject analysis, we showed that trial - to - trial variation in low - Moreover, in a within level feature reactivation predicted greater recognition accuracy, albeit onl y for participants with higher - than - average recognition performance on lure trials. Episodic Memory - Feature Specific Neural Reactivation during Using FSIC, we found that visual features from all selected levels of the CNN were represented, to some degree, throughout the cortical visual hierarchy; but these representations were not evenly distributed across ROIs (see the diagonal of Figure 4b). Consistent with previous work indicating a correspondence between the hierarchical organization of the layers of a CNN and the Horikawa & ; 2015 Güçlü & van Gerven, cortical regions of the visual processing stream ( 2017 ), the distribution of features, revealed by FSIC and the locations of peak neural Kamitani, - to ed according to the posterior level (Figure 4a), was organiz - reactivation for each feature -

21 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . anterior visual hierarch y. This organization was also evident in the correlations between cortical neural reactivation, imagery vividness ratings and recognition response accuracy (Figure 5), y to use CNN feature vectors to more precisely characterize and establishing the abilit thereby decompose the content of episodic memories. - Consistent with prior work we found that the association between feature levels from the - CNN and neural activity was largely congruent with the well known hierarchical organization of level visual features - er evidence for low strong found the visual system. However, we also - - order represented within cortical regions , and higher - level features within lower order higher feed - regions (Figure 4b ). Unlike strictl the y forward CNN’s, like the one used in this study, back connections that can - forward and feed - a complex network of both feed comprises cortex lower - and higher bypass intermediate areas, order facilitating direct communication between - Hegde Felleman , 2007 ; regions ( Desi mone et al., 1984; Lamme, Super & Spekreijse, 1998 ; & and ion modulat maintenance, combination of features enabling the , thereby et al., 2013) Kravitz & Mangun 2000 at multiple levels Gilbert & ( Chun & Jiang , 1999 ; Hopfinger, Buonocore ; & Gazzaley, 2010 Zanto, Rubens, Thangavel, & Sigman, 2007 ; Zanto, Rubens, Bollinger ; ) For example, the Piëch, Li, Reeke ; Gazzaley & Nobre, 2012 ; 2011 . Gazzaley, & Gilbert, 2013 selective maintenan relevant visual - ce of task has been implicated in the s inferior frontal gyru down connections with the visual cortex during working memory and mental - information via top Mayer et al., 2007; Zanto, Rubens, imagery ( Vandenberghe et al., 1996 ; Nobre et al., 2004; Bollinger & Gazzaley, 2010 ; Higo et al., 2011; Dijkstra, Zeidman, Ondobaka, Gerven, & Friston, Through the use of FSIC, w e show that the inferior frontal gyrus contains representations ). 2017 of naturalistic sce , of visual features from all levels of the visual hierarchy during the nes recall with the and that the reinstatement of these representations within the IFG are correlated

22 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . reinstatement of the same features within the occipital cortex, supporting the idea that the region . sual areas in early vi specific neural reactivation - facilitates feature level - level visual Low , high - level visual , and semantic features, but not mid - visual inferior order visual and frontal regions beyond the - features, were identified in many higher ( level and - . While this was expected of high frontal gyrus semantic features the features because the frontoparietal cortex level features within - , the prevalence of low represent object identity) and higher . more surprising This raises the order regions of the ventral visual stream - was order regions — question of the function of such low - level features within these putative high er - a question that recent advances within the field of computational neural networks may shed some liams, light upon. Like the receptive fields of neurons within the visual (Smith, Singh, Wil cortex feedforward comprise the nodes that , & Greenlee, 2001; Rolls, Aggelopoulos, & Zheng, 2003) are organized such that CNNs designed to perform visual classification and localization tasks the lower - order layers have small receptive fields and weak sema ntics , whereas the higher - order and strong semantics layers have large receptive field s ( Luo, Li, Urtasun, & Zemel, 2016 ). Consequently, the of the semantic - sensitive layers is low , resulting in the loss of fine resolution details essential for some tasks (e.g. the classification of small objects) . To address this problem, d and “skip” connections (which - down connections more recent CNNs have incorporate top - order layers of the bypass adjacent layers) to directly combine the outputs of lower - and higher et al., network , thereby increasing the effective resolution of the semantic - sensitive layers ( Liu . 2018) This approach has been proven to be effective for a variety of tasks requiring both ng classification and localization of small accurate semantics and fine visual details, includi a key objects (Shrivastava, Sukthankar, Malik, & Gupta, 2016 ), and salient object detection ( attention al processes ) and boundary delineation element of the coordination of (important for

23 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . et al., 2017). Given the functional roles of the grasping behavior , among other tasks ) ( Zhang ) Spector, & Weiner, 2014 - Grill order ventral visual stream in visual object classification ( - higher Schn ider, & and the frontoparietal cortex in attention and grasping (Ptak, 2012; behavior Ptak, visual representations within level - he presence of low e posit that t ), w 2017 Fellrath, these allocation and motor planning regions , attentional may likewise facilitate visual classification more s te semantics and fine visual detail specifically, and any task that requires both accura tasks generally. Neural Reactivation and Memory Vividness Feature - Specific T o investigate the functional contributions of feature - specific neural reactivation to memory , we hypothesis that the vividness of test ed the memory recall should positively correlate with the - . degree of neural reactivation encoding visual features — particularly low level visual features found l reactivation Although previous research correlation s d ha between vividness and neura (Cui et al. 2007; throughout early and late regions of the ventral and dorsal visual streams Johnson et al. 2015; St - Laurent, Abdi, and Buchsbaum 2015; Dijkstra, Bosch and van Gerven - 2017; Bone et al., 201 9 ), the relative contributions of th e reinstatement of lower - and higher features from level visual features remained an open question. B y measuring the reactivation of evel based upon the different levels of the visual hierarchy , as opposed to inferring feature - l , we f ound the reinstatement location of reactivation (i. ( Poldrack, 2011 e. reverse inference ) ) that to an approx imately equal of low er - and higher - level visual features correlated with vividness vividness should correlate with degree . While our hypothesis did predict that reinstatement of the low - level correlation was expected to be stronger - and high - level visual features both low , is a vivid memory constituting of visual details recall based upon the assumption that the

24 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . (Hebb, 1968; Kosslyn, Ganis, primarily the reinstatement of low - l evel features dependent upon ). & Thompson, 2001 , however, may overlook This assumption from high - the inference of low level features - level features . According to the predictive coding account of perception , visual experience cortical reciprocal exchange of bottom - - down up and top signals throughout the results from the , 20 10; Bastos et al., 2012 ) . Top - down signals hierarchy ( Rao & Ballard, 1999 ; Friston, 2005 h from neurons representing igh - level features, which encode statistically and behaviorally serve to drive and/or modulate neural significant non - linear combinations of lower level features, functioning as a generative representing activity the associated lower - level features thereby — down - these top model of how environmental stimuli cause sensations . During perception, connections convey predictions, which are compared against the perceptual input to generate an . This signal is then propagated back up the hierarchy to upd ate the predictions (i.e. error signal alter higher - order activations) and enhance memory of the features that diverged from Axmacher uring episodic memory ( et al., 2010 ; Henson & Gagnepain, 2010 ). D expectations sparsely - r level features are used to infer lowe - higher cued , recall level features, while the lower level features recalled that were not accurately predicted during perception serve to to be specific to the recalled episode Therefore . according this inference constrain , predictive to a visual details (i.e. coding account of visual recall , the number and accuracy of remembered and low of both high level features . memory vividness) should depend upon the reactivation - - Moreover, because participants were instructed to not rate generic imagery related to the cue as level features that were not present in the encoded image vivid, the top - down inference of low - should correlate negatively with vividness , which is what we found. Thus , the partial correlations a are consistent with specific neural reinstatement - between subjective vividness and feature

25 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . memory recall. predictive coding account of visual perception and Specific Neural Reactivation and Recognition Accuracy - Feature connection between feature demonstrate a - serve to our vividness results Whereas specific neural the establishing we were also interested in of memory, quality subjective reactivation and the memory : relationship between neural reactivation and an objective memory measure recognition memory access to required task participants performed for our study recognition The . accuracy grained memory information in order to identify a probe image drawn from the same - fine old or new . Given the strong semantic semantic category (e.g. two images of a steam train) as like features overlap between the two images, higher - level semantic - alone would be unlikely to we hypothesized that the provide enough information to distinguish the images. Consequently, recall would be required to s feature level - lower of perform well on the task. Overall, our results (Figure 5d and Supplementary Figure 4 d) . We found that reactivation supported this hypothesis - of lower level features within the early visual cortex positively correlated with recognitio n - subjects, albeit the within - subject result only held for subjects accuracy within - and between - . on lure trials average recognition accuracy - with greater than What might be the cause of this individual difference in the relationship between neural ation and One possibility is that the participants differ in their reliance accuracy? recall reactiv upon the reinstatement of higher - vs. lower - level features when comparing the presented image with the memorized image . Our original hypothesis that reactivation of lower - level features all subjects subjects assumed should positively correlate with recognition accuracy within - that level representations when performing the task. Our failure to find the - would utilize lower appears to be the result of greater than expected individual subject - pothesized within hy effect

26 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . variation in the ability or tendency of subjects to reactivate low - level visual features during these ause and implications of the c will be required to explore Future studies . memory retrieval s important . individual difference Limitations are reactivated We have demonstrated that features at all levels of the visual hierarchy - throughout the cortex during such shown how Moreover, we have episodic recall. feature spe reactivation relates to the vividness of cific To , and subsequent recognition accuracy. recall obtain these findings, it was necessary to develop and utilize methods that control for the inherent correlations r simulation results strongly between feature - levels (e.g. FSIC). Although ou with As indicate that our methods were successful in this regard, a caveat must be considered. VGG16 ( the any model of feature - specific cortical representations, t he features extracted from pected to be C NN used in this study ) cannot be ex a complete set of all visual features can not represented within a given participant’s cortical activity. Consequently, our approach - ing in the false detection of potentially level correlations, for all inter control exhaustively result - feature - specific neural reactivation . To address this concern, consider t wo hypotheticals : lower - level features tend to be detected within regions that only contain higher level features, and /or - tend to be d etected within regions that only contain lower - level features. If higher level features - level features the former was true, we would expect approximately equal reactivation of mid regions , due to low - level and low - relative to - level features within high er - order cortical the mid to approximately the same degree level features correlati ng with higher - level feature s reactivation Supplementary Figure 1 was much more pronounced than ( ) . In contrast, low - level reactivation ) ; first two rows along the diagonal (Figure 4b latter was true, then level - mid . If the

27 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . semantic features would be expected within the earliest region of the visual cortex: the calcarine Therefore, our results ). ; last row along the diagonal sulcus. This was not the case (Figure 4b FSIC, successfully strongly indicate that the methods used e.g. within the current study, controlled for the correlations between features drawn from different levels of the visual , thereby eliminating the false positives (e.g. that previous approaches hierarchy Horikawa & were susceptible to . 2017 Kamitani, ) Conclusion The contributions of this study were fourfold. First, we developed novel measures of feature - specific neural reactivation, e.g. FSIC, that control for the inherent correlations between s without sacrificing sensitivity. Second, the results - hierarchically organized feature level obtained from FSIC revealed that neural reactivation during episodic memory is more particularly for low - level features (e.g. edges) — which we widespread than previously thought — posit subserves a multit ude of cognitive functions requiring both fine visual detail and accurate object/scene categorization (e.g. fine grasping behavior). Third, we found that neural reactivation level visual features contributed equally to the subject of lower - level and higher - ive vividness of recall, which we argue supports a predictive coding account of perception and recall. Lastly, we confirmed that when differentiating semantically nearly - identical images from memory only with recognition accuracy. Overall, the reactivation of low - level visual features correlates - multifaceted nature of the current study’s results shows the potential for FSIC, and other feature specific approaches to decomposing neural pattern representations, to test and elucidate the neural underpinnings of memory and mechanisms underpinning long held theories about the cognition.

28 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . Materials and Methods Participants - right seven - Thirty - normal vision and no history to handed young adults with normal or corrected - of neurological or psychiatric disease were recru ited through the Baycrest subject pool, tested and paid for their participation per a protocol approved by the Rotman Research Institute’s Ethics Board. Subjects were either native or fluent English speakers and had no contraindications for MRI. Data from ten of these participants were excluded from the final analyses for the 7 ), ( 3 did not complete experiment ). Thus, twenty - following reasons: excessive head motion ( old seven participants were included in the final analysis ( 1 5 males and 12 females, 20 - 32 years [mean: 25 ]). Stimuli (800 by 600) were gathered from online sources . colored photographs 111 For each image, an image 111 image pair was acquired using Google’s similar image search function, for a total of were used pairs ( for practice, and the remaining 90 were used 222 images). 21 image pairs during the in - scan encoding and retrieval tasks ( see Supplementary Figure 5 for example image audio pairs ). Each image was paired with a short descriptive title in a synthesized female voice ; this title served as a retrieval cue runs ) during encoding ( https://neospeech.com ; voice: Kate ( 720 by 480 pixels; 10m during the in - scan retrieval task . Two videos used for model training composed 25s and 10m 35s in length ) were of a series of short (~4s) clips drawn from YouTube themes (e.g. still photos of bugs, people performing and Vimeo , containing a wide variety of Indiana Jones: Raiders of the “ video cut from manual tasks, animated text, etc.). One additional

29 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . Ark ” (1024 by 435 pixels; 10m 6s in length) was displayed while in the scanner, but the Lost associated data was not used in this experiment. Procedure Before undergoing MRI, participants were trained on a practice version of the task incorporating three video viewing runs 21 practice image pairs MRI . Inside the scanner, participants completed encoding three and - retrieval sets . The order of the runs was a s follows: first video viewing run 2 1 ), third video viewing run (Indiana Jones ), second video viewing run (short clips (short clips first encoding - retrieval set, second encoding - retrieval set, third encoding - retrieval set . A clip), encoding retrieval third and - second resolution structu - high ral scan was acquired between the sets , provid ing a break. w s Video viewing runs were 10 m 57s seconds long. For each run, participant ere instructed to pay attention while the video (with audio) played within the center of the screen. The order of the videos was the same for all participants. followed by retrieval sets were composed of one encoding run - Encoding one retrieval required the participants to first memorize and then recall 30 images drawn from 30 run. Each set image pairs The image pairs within each set were selected randomly, with the constraint that no . image pair to be The image selected from each image pair could be used in more than one set. This experimental procedure presented during encoding was counterbalanced across subjects. was designed to limit the concurrent memory load to 30 images . 24 s long. Each run started with 10s during which instructions Encoding runs were 6 m s the appearance of image in the center of the with began an rial T screen. - were displayed on

30 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . s) , accompanied by a simultaneous descriptive audio cue (e.g. a picture depicting screen ( . 8 1 600 by 800 pied . Images occu toddlers would be coupled with the spoken word “toddlers”) pixels of a 1024 by 768 pixel screen. Between trials, a cross - hair appeared in the center of the screen to . Participants were instructed to pay attention to each image and 7s (font size = 50) for 1. they could visualize the images as precisely as encode as many details as possible so that possible during the task. imagery To assess ongoing engagement with the task, the participants back task requiring the participants to press “1” if the displayed image was - 1 also performed a the same as the - back task stimuli for the 1 Within each run, preceding image, and “2” otherwise. were randomly sampled with the following constraints: 1) each image was repeated exactly four one immediate times in the run (120 trials per run; 360 for the entire scan), 2) there was only back - repetition per image, and 3) the other two repetitions were at least 4 items apart in the 1 sequence. runs were s long Retrieval 32 . Each run started with 10s during which instructions m 9 n cued once each (the order was randomized), were displayed on - screen. Thirty images were the for a total of 30 trials per run (90 for the entire scan). T rial s began with an image title appeared in the center of the screen for 1s (font = Courier New, font size = 30). After 1s, the title was 6 replaced by a n empty rectangular box shown in the center of the screen ( s), and whose edges 800 pixels). Participants were 600 corresponded to the edges of the stimulus images ( by much detail instructed to visualize the image that corresponded to the title as accurately and in as as they could within the confines of the box. Once the box disappeared, participants were prompted to rate the vividness visual details specific recalled (defined as the relative number of 3 s) using a to the cued image presented during encoding ) of their mental image on a 1 - 4 scale ( right 4 = right little finger). index finger; ; 1 = right hand - four button fiber optic response box (

31 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . image (800 by 600 pixels) in th e center of the This was followed by the appearance of a probe the image either or similar to the trial’s cued image (i.e. the same as screen (3s), that was either participants shown during encoding or its pair). While the image remained on the screen, the were instructed to respond with “1” if they th ought that the image was the one seen during encoding , or “2” if the image was new (responses made using the response box). Following (old) the disappearance of the image, participants were prompted to rate their confidence in their old/new response on a 1 hair Between each trial, a cross 4 scale (2s) using the response box. - - . 1 either (font size = 50) appeared in the center of the screen for , 2 or 3 seconds both images within each image pair Randomization sequences were generated such that . During (image A and B) w ere presented equally often during the encoding runs across subjects > probe A) or - retrieval runs each image appeared equally often as a matching (encode A > probe B) image across subjects. Due to the need to remove several - mismatching (encode A subjects from the analyses, stimulus versions were approximately balanced over subjects. Setup and Data Acquisition - Participants were scanned with a 3.0 - T Siemens MAGNETOM Trio MRI scanner using a 32 weighted scan coplanar - o multi - slice T1 - channel head coil system. A high - resolution gradient ech - planar imaging scans (EPIs) was first acquired for localization. Functional images with the echo were acquired using a multiband EPI sequence sensitive to BOLD contrast (22 x 22 cm field of plane resolution of 2 x 2 mm for each of 63 - ix size, resulting in an in view with a 110 x 110 matr 2 mm axial slices; repetition time = 1.77 sec; echo time = ms; flip angle = 62 degrees). A 30 - - resolution whole - brain magnetization prepared rapid gradient echo (MP - RAGE) 3 - high D T1 ghted scan (160 slices of 1mm thickness, 19.2 x 25.6 cm field of view) was also acquired for wei

32 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . anatomical localization. The experiment was programmed with Experiment Builder version 1.10.1025 (SR ere projected onto a screen Research Ltd., Mississauga, Ontario, Canada). Visual stimuli w behind the scanner made visible to the participant through a mirror mounted on the head coil. fMRI Preprocessing 1 format, motion Functional images were converted into NIFTI - - corrected and realigned to the 3dvolreg program. The maximum average image of the fi rst run with AFNI’s (Cox 1996) displacement for each EPI image relative to the reference image was recorded and assessed for weighted - - registered to the high - head motion. The average EPI image was then co resolution T1 MP - RAGE structural using the AFNI program align_epi_anat.py ( Saad et al, 2009) . back The volumetric functional data for each experimental task (video viewing, 1 - specific cortical surface generated - encoding task, retrieval task) was then projected to a subject ( Dale, Fischl & Sereno, by F ) . The target surface was a spherically normalized reesurfer 5.3 1999 mesh with 32000 vertices that was standardized using the resampling procedure implemented in ) the AFNI program MapIcosahedron ( Argall, Saad & Beauchamp, 2006 . To project volumetric imaging data to the corti cal surface we used the AFNI program 3dVol2Surf with the “average” mapping algorithm, which approximates the value at each surface vertex as the average value among the set of voxels that intersect a line along the surface normal connecting the white matte r and pial surfaces. The three video scans (experimental runs 1 3), because they involved a continuous - processing to the - stimulation paradigm, were directly mapped to the surface without any pre

33 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . cortical surface. The three retrieval scans (runs 5, 7, 9) were first divided into a sequence of 2) two seconds prior to the onset of the retrieval - experimental trials with each trial beginning (t= cue (verbal label) and ending 32 seconds later in two second increments. These trials were then series, each of which consisted of specific time - trial concatenated in time to form a series of 90 - - wise data blocks were then projected onto the cortical surface. To 16 samples. The resulting trial facilitate separate analyses of the “imagery” and “old/new ” retrieval data, a regression judgment expected hemodynamic response associated with the appr oach was implemented. For each trial, each task was generated by convolving a series of instantaneous impulses over the task period ic response. (10 per second; imagery: 61; old/new: 31) with the SPM canonical hemodynam Estimates of beta coefficients for each trial and task were computed via a separate linear per trial (each with 16 samples: one per time point), with vertex activity as the regression dependent variable, and the expected hemodynamic response values for the “imagery” and Finally, data from the three encoding scans (runs “old/new judgment” as independent variables. 4, 6, 8) were first analyzed in volumetric space using a trial - wise regression approach, where the onset of each image stimulus was modelled with a separate regressor formed from a convolution instantaneous impulse with the SPM canonical hemodynamic response. Estimates of trial of the - wise beta coefficients were then computed using the “least squares sum” ( Mumford, Turner, 2012 ) regularized regression approach as implemented in the AFNI program Ashby & Pold rack, 3dLSS . The 360 (30 unique images per run, 4 repetitions per run, 3 total runs) estimated beta 3dVol2Surf. coefficients were then projected onto the cortical surface with ural Network Image Features Deep Ne We used the implementation of the VGG16 deep neural network (DNN) model TensorFlow

34 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . for the ( & Zisserman, 2014 ; see http://www.cs.toronto.edu/~frossard/post/vgg16 Simonyan The VGG16 model consists of a total of thirteen convolutional layers and ). implementation used video frames ( 90 three fully connected layers. 3 from the memory task and image pairs 3775 o 1: 1875 frames; video 2: 1900 clip videos; vide - short per second; taken from the two frames the VGG1 frames model for each of 6 ) were resized to 224 × 224 pixels to compute outputs the seventh the second convolutional layer (layer 2), . The outputs from the units in image/frame convolutional layer (layer 7), fully connected convolutional layer (layer 13), and the the last final mid - level were treated as vector s corresponding to low - layer (layer 16) level visual features, visual features, high - level visual features and semantic features, respectively Layer selection . 1c w as performed manually by inspection of filter activations (see Figure for example T o account for the low retinotopic spatial resolution resulting from participants eye activations). movements, the spatial resolution of the convolutional layers (the full y connected layer has no explicit spatial representation) was reduced to 3 by 3 (original resolution for layer 2: 224 by 224; level visual features, layer 7: 56 by 56; layer 13: 14 by 14). The resultant vector length of low - level visual features, high - level visual features and semantic features was 576, 2304, 4608 mid - improve to transformed - Convolutional layer activations were log and 1000, respectively. (Naselaris prediction accuracy . et al., 2015) Encoding Model feature levels and each individual voxel four of the each An encoding model was estimated for during trial t . The encoding model for et al., 2015). Let v (Naselaris be the signal from voxel i it this voxel is: 푇 푓 + = 푣 휖 ℎ 푖푡 푡

35 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . associated with the current trial/image Here is a 100 × 1 vector of 100 image features f t (only the 100 features with the largest positive correlations with voxel activity were selected to vector of model parameters that indicate the h , make the computation tractable) a is 100 × 1 - and ε is zero indicates transposition) T (the superscript voxel’s sensitivity to a particular feature mean Gaussian additive noise. (R package - h were fit using non The model parameters regression negative lasso “nnlasso”; Mandal , 2016 ) trained on data drawn from the encoding and movie viewing & Ma fold cross validation over the encoding data using 3 tasks (excluding the Indiana Jones video) - negative constraint was included to reduce the (all movie data was used in each fold). The non - lity that a complex linear combination of low possibi - level features may approximate one or more 5 log - . (lambda) The regularization parameter was determined by testing high s level feature - . For path feature) 0 to 1 (using the nnlasso function’s 1 approximately spaced values from /1000 the model parameters h were estimated for each voxel each value of the regularization parameter , was measured and then prediction accuracy (sum of squared errors; SSE) of the re cognition data the model parameters , . For each voxel fold cross validation - 3 using that yielded the highest h prediction accuracy were retained for . image decoding Decoding Image s for each unique combination w ncoding model E ere used to predict neural activity recall during 74 bilateral cortical FreeSurfer ROIs ) . The of subject, feature - level , ROI , and retrieval trial ( accuracy of this prediction was assessed as follows: 1) f or each combination of subject, feature - for the 90 images viewed during the ed neural activation patterns level, and ROI t he predict task trained on the movie and encoding using a model that was generated were ng task encodi

36 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . data using 3 - fold , excluding data from encoding trials wherein the predicted image was viewed w (across vertices within s prediction correlate for each retrieval trial, the . 2) cross validation ere d resulting in 90 correlation recall during with the observed neural activity ) the given ROI he correlation coefficients ranked in descending order , and the rank of the coefficients were . 3) t recorded (1 = highest accuracy, 90 = lowest was image recalled prediction associated with the his rank was chance, and a subtracted from t accuracy). 4) was then mean rank (45.5) so that 0 the (44.5 = highest positive value indicated greater - than - chance accuracy for the given trial - accuracy, 44.5 = lowest accuracy) . Seed ROI Selection S The . s feature level of the four for each selected from all Freesurfer ROIs eparate ROIs were the p (Figure 3a) was as follows rocedure for generating weight values for each ROI 1) g et : cation accuracy across subjects during image perception ( data taken from the average classifi level during the retrieval z - score old/new recognition task blocks ) for each feature - and ROI. 2) values less et s . 3) level - for each feature classification accuracy across ROIs all than zero to zero. level - associated with the subtract the greatest value , other feature - feature ROI and or each f 4) from the target feature’s value . 5) s et all values less than zero to zero. 6 ) levels ormalize the n i.e. divide each value by the sum of all values ) for each values across ROIs to sum up to one ( et values less than .05 to 0 feature level . 7 ) s to retain only those ROIs that have a strong - . 8 ) n ormalize the values across ROIs for each feature level. association with the feature level pecific Informational Connectivity S - Feature during FSIC For the analyses, correlations image classification accuracy (rank measure) of

37 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . for all combinations imagery were performed across ROIs. orrelations were performed Separate c Selection”) and all Seed ROI esurfer ROIs as outlined in “ seeds (selected from all Fre four of was drawn from other ROIs not included within the seed. For each seed, classification accuracy linear mixed the seed’s associated feature level . The correlations were calculated with a - effects for the seed classification accuracy trials, wherein (LME) model on data from all episodic recall four feature levels ROI was the dependent variable (DV), classification accuracy for each of the ticipant and image were within the target independent variables (IV), and par ROI were the crossed random effects (random intercept only, due to model complexity limitations). Statistical - , calculated with the BootMer function assessments were performed using bootstrap analyses rrected for multiple comparisons across ROIs using (Bates et al. 2015) using 1000 samples and co . FDR ( Benjamini & Hochberg, 1995 ) fMRI Data Simulation The simulation used the same experimental structure and stimuli (for training and testing the For each simulated subject, 800 artificial voxels were created, models) as the true experiment. ed from the CNN VGG16 as feature extract , randomly selected, with each voxel containing one - described in the “Deep Neural Network Image Features” section. For each voxel, the feature - specific activation associated with the video frame or image presented at each time point or trial oxels were grouped into 8 ROIs with 100 voxels V . y was used to simulate the voxel’s activit There were 2 ROIs per feature - level (one representing the seed ROI, and the other each. representing the target ROI), such that features assigned to the voxels in each ROI were extracted level. The two ROIs assigned the same level contained identical features, i.e. from the assigned . For gaussian noise of independent except for the subsequent application duplicates, they were

38 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . 2b , the seed ROIs contained 25 voxel s the analysis depicted in Figure Supplementary was recall during Memory loss . levels (for 100 voxels total) - representing each of the four feature by randomly setting a fraction of the features to zero. The same features were set to simulated ROI - imulating cross zero across ROIs given trial, s for a representing the same feature level - information transfer Trial trial variation in memory accuracy was simulated by varying the by . - - fraction of feature loss over trials (randomly selected using a uniform distribution from 40 an was added to all data, (mean 0, standard deviation 1) noise Lastly, independent gaussi 95%). ), to , equally distributed with the SNR varying across simulated subjects (either 15, 25 or 35% simulate all unaccounted - for variation in voxel activity, and individual variations thereof. Linear Models and Statistics ) was performed using a and 3a Statistical assessment of mean neural reactivation (Figure 2 ROI, with neural reactivation as the DV and subject and image as separate LME model for each - values were calculated with bootstrap crossed random effects. Confidence interval s and p statistical analyses (1000 samples) using the BootMer function (Bates et al. 2015) and corrected for multiple comparisons across ROIs using false discovery rate ( FDR; Benjamini & Hochberg, - vividness ratings 1995 correlations between feature - specific reactivation , ). For the within subject , LME models or accuracy recognition (Figure 5b) vividness ratings and were used, with , the four neural as the dependent variable (DV), recognition accuracy (correct vs incorrect) - level and higher - level) and feature - reactivation measures for each combination of ROI (lower level) level (lower - level and higher as independent variables (IV), and participant and image as - - to model complexity limitations). intercept only, due crossed random effects (random values were calculated with bootstrap statistical analyses (1000 - Confidence intervals and p

39 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . samples) using the BootMer function and corrected for multiple comparisons across coefficients specific reactivation and - correlations between feature subject - For the between using FDR. was used, with recognition accuracy ), a accuracy (Figure 5 c single linear model recognition (correct vs incorrect) as the dependent variable (DV) and the four neural reactivation measures as and p values - were generated with bootstrap nt variables (IV). independe Confidence intervals statistical analyses (1000 samples) across using FDR and corrected for multiple comparisons — including the four coefficients from the within - subject recognition accuracy LME coefficients (i.e. eight coefficients in total) . Funding This work was supported by the Natural Sciences and Engineering Research Council of Canada and the Canadian Institutes of Health Research (152879 to B.R.B) . ) ( 488937 to B.R.B. Author Contributions Co nceptualization, M.B.B., B.R.B.; Methodology, M.B.B., B.R.B., F.A.; Software, M.B.B., B.R.B.; Formal Analysis, M.B.B.; Investigation, F.A.; Data Curation, M.B.B., B.R.B.; Writing – Original Draft, M.B.B.; Writing – Review and Editing, M.B.B., B.R.B.; Visua lization, M.B.B.; Supervision, B.R.B. Conflicts of Interest There are no competing interests.

40 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . References Albers, A. M., Kok, P., Toni, I., Dijkerman, H. C., & de Lange, F. P. (2013). Shared ual cortex. Current representations for working memory and mental imagery in early vis 1431. - Biology, 23(15), 1427 Anzellotti, S., & Coutanche, M. N. (2018). Beyond functional connectivity: investigating - 269. networks of multivariate representations. Trends in cognitive sciences, 22(3), 258 Argall, B. D., Saad, Z. S., & Beauchamp, M. S. (2006). Simplified intersubject averaging on the - 27. cortical surface using SUMA. Human brain mapping, 27(1), 14 Axmacher, N., Cohen, M. X., Fell, J., Haupt, S., Dümpelmann, M., Elger, C. E., Schlaepfer, T. , Lenartz, D., Sturm, V. & Ranganath, C. (2010). Intracranial EEG correlates of E. expectancy and memory formation in the human hippocampus and nucleus accumbens. Neuron, 65(4), 541 549. - ropsychological Baddeley, A. D. (1988). Imagery and working memory. In Cognitive and neu 180). Springer, Dordrecht. approaches to mental imagery (pp. 169 - Bastos, A. M., Usrey, W. M., Adams, R. A., Mangun, G. R., Fries, P., & Friston, K. J. (2012). - Canonical microcircuits for predictive coding. Neuron, 76(4), 695 711. M . , Bolker , B . , Walker , S. ( 2015 ) . Fitting linear mixed - effects models using , D . , Maechler Bates , 48. - lme4. Journal of Statistical Software. 67:1 Benjamini, Y., & Hochberg, Y. (1995). Controlling the false discovery rate: a practical and e testing. Journal of the Royal statistical society: series B powerful approach to multipl

41 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . (Methodological), 57(1), 289 300. - - Bichot, N. P., Heard, M. T., DeGennaro, E. M., & Desimone, R. (2015). A source for feature 844. - based attention in the prefrontal cortex. Neuron, 88(4), 832 ley, J. K., Jones, G. P., & Town, S. M. (2016). Where are multisensory signals combined for Biz - 37. making?. Current opinion in neurobiology, 40, 31 perceptual decision - Laurent, M., Dang, C., McQuiggan, D. A., Ryan, J. D., & Buchsbaum, B. R. Bone, M. B., St - Eye - movement reinstatement and neural reactivation during mental imagery. ). 9 01 (2 Cerebral Cortex, 29 3 ), 1075 – 1089 . ( Rodger, S., Fang, C., Abdi, H. (2012). The neural basis of vivid Buchsbaum, B. R., Lemire - memory is patterned on perception. Journal of Cognitive N euroscience. 24 :1867 - 1883. down versus bottom Buschman, T. J. & Miller, E. K. (2007). Top - - up control of attention in the 1862 prefrontal and posterior parietal cortices. Science. 315, 1860 – . Cabeza, R., Ritchey, M., Wing, E. A. (2015). Reinstatement of Individual Past Events Revealed by the Similarity of Distributed Activation Patterns during Encoding and Retrieval. 691. 27:679 Journal of Cognitive N euroscience. - Carota, F., Kriegeskorte, N., Nili, H., & Pulvermüller, F. (2017). Represent ational similarity mapping of distributional semantics in left inferior frontal, middle temporal, and motor - 309. cortex. Cerebral Cortex, 27(1), 294 down attentional guidance based on implicit learning of - Chun, M. M., & Jiang, Y. (1999). Top 365. - visual covaria tion. Psychological Science, 10(4), 360

42 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . Cichy, R. M., Heinzle, J., & Haynes, J. D. (2011). Imagery and perception share cortical 380. - representations of content and location. Cerebral Cortex, 22(2), 372 ). Informational connectivity: identifying Schill, S. L. (2013 - Coutanche, M. N., & Thompson - voxel patterns across the brain. Frontiers in human synchronized discriminability of multi neuroscience, 7, 15. Cui, X., Jeter C. B., Yang D., Montague P. R., Eagleman D. M. (2007). Vividness of mental imagery: in dividual variability can be measured objectively. Vision Research. 47:474 - 478. - Dale, A. M., Fischl, B., & Sereno, M. I. (1999). Cortical surface based analysis: I. Segmentation - and surface reconstruction. Neuroimage, 9(2), 179 194. J. R. (2010). The ghosts of brain states past: remembering reactivates Danker, J. F., Anderson, 136: the brain regions engaged during encoding. Psychological Bulletin. 87. Dentico, D., Cheung, B. L., Chang, J. Y., Guokas, J., Boly, M., Tononi, G., & Van Veen, B. (2014). Reversal o f cortical information flow during visual imagery as compared to , visual perception. Neuroimage 100 , 237 - 243. Desimone, R., Albright, T. D., Gross, C. G., & Bruce, C. (1984). Stimulus - selective properties of 2062. - l of Neuroscience, 4(8), 2051 inferior temporal neurons in the macaque. Journa Destrieux, C., Fischl, B., Dale, A., & Halgren, E. (2010). Automatic parcellation of human cortical gyri and sulci using standard anatomical nomenclature. Neuroimage, 53(1), 1 - 15. . A. (2017). Vividness of visual imagery depends on the Dijkstra, N., Bosch, S., & van Gerven, M

43 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . neural overlap with perception in visual areas. Journal of Neuroscience, 3022 16. - - Dijkstra, N., Zeidman, P., Ondobaka, S., Gerven, M. A. J., & Friston, K. (2017). Distinct top connectivity during visual perception and imagery. Scientific up brain - down and bottom reports, 7(1), 5677. Eickenberg, M., Gramfort, A., Varoquaux, G., & Thirion, B. (2017). Seeing it all: Convolutional 184 - 194. network layers map the function of the human visual system. NeuroImage, 152, Ester, E. F., Sprague, T. C., & Serences, J. T. (2015). Parietal and frontal cortex encode stimulus - - specific mnemonic representations during visual working memory. Neuron, 87(4), 893 905. K. (2001). Categorical representation Freedman, D. J., Riesenhuber, M., Poggio, T., & Miller, E. 316. - of visual stimuli in the primate prefrontal cortex. Science, 291(5502), 312 Friston, K. (2005). A theory of cortical responses. Philosophical Transactions of the Royal Society of London B: Biological Sciences, 36 0(1456), 815 - 836. Friston, K. (2010). The free - energy principle: a unified brain theory?. Nature reviews neuroscience, 11(2), 127. Gazzaley, A., & Nobre, A. C. (2012). Top - down modulation: bridging selective attention and 135. e sciences, 16(2), 129 - working memory. Trends in cognitiv down influences in sensory processing. - Gilbert, C. D., & Sigman, M. (2007). Brain states: top 696. - Neuron, 54(5), 677 Spector, K., & Weiner, K. S. (2014). The functional architecture of the ventral temporal - Grill

44 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . cortex and its role in categorization. Nature Reviews Neuroscience, 15(8), 536. Güçlü, U., & van Gerven, M. A. (2015). Deep neural networks reveal a gradient in the complexity of neural representations across the ventral stream. Journal of Neuroscience, 10014. - 10005 35(27), Hebb, D. O. (1968). Concerning imagery. Psychological Review. 75:466. Hegde, J., & Felleman, D. J. (2007). Reappraising the functional implications of the primate visual anatomical hierarchy. The Neuroscientist, 13(5), 416 - 421. & Gagnepain, P. (2010). Predictive, interactive multiple memory systems. Henson, R. N., 1326. - Hippocampus, 20(11), 1315 Harrison, S. A., & Tong, F. (2009). Decoding reveals the contents of visual working memory in early visual areas. Nature, 458(7238), 632. Hesslow, G. (2012). The current status of the simulation theory of cognition. Brain Research. - 9. 1428:71 Higo, T., Mars, R. B., Boorman, E. D., Buch, E. R., & Rushworth, M. F. (2011). Distributed and roceedings of the National causal influence of frontal operculum in task control. P Academy of Sciences, 108(10), 4230 - 4235. Hopfinger, J. B., Buonocore, M. H., & Mangun, G. R. (2000). The neural mechanisms of top - down attentional control. Nature neuroscience, 3(3), 284. ric decoding of seen and imagined objects using Horikawa, T., & Kamitani, Y. (2017). Gene hierarchical visual features. Nature communications, 8, 15037.

45 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . Hung, C. P., Kreiman, G., Poggio, T., & DiCarlo, J. J. (2005). Fast readout of object identity 866. - 0(5749), 863 from macaque inferior temporal cortex. Science, 31 Huth, A. G., Nishimoto, S., Vu, A. T., & Gallant, J. L. (2012). A continuous semantic space describes the representation of thousands of object and action categories across the 1224. - human brain. Neuron, 76(6), 1210 ., Ungerleider, L. G. (2002). Visual imagery of famous faces: effects of Ishai, A., Haxby, J. V 41. - memory and attention revealed by fMRI. Neuroimage. 17:1729 Johansson, R., Holsanova, J., Dewhurst, R., Holmqvist, K. (2012). Eye movements during scene nal role, but they are not reinstatements of those produced recollection have a functio during encoding. Journal of Experimental Psychology: Human Perception and Performance. 38:1289. Johnson, M. R., Johnson, M. K. (2014). Decoding individual natural scene representations during perc eption and imagery. Frontiers in Human Neuroscience. 8:59. Johnson, M. K., Kuhl, B. A., Mitchell, K. J., Ankudowich, E., Durbin, K. A. (2015). Age - related differences in the neural basis of the subjective vividness of memories: Evidence from multivoxel pat tern classification. Cognitive, Affective, & Behavioral Neuroscience. 61. 15:644 - Kay, K. N., Naselaris, T., Prenger, R. J., & Gallant, J. L. (2008). Identifying natural images from human brain activity. Nature, 452(7185), 352. 4). The sensory strength of voluntary visual imagery predicts Keogh, R., & Pearson, J. (201

46 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . visual working memory capacity. Journal of vision, 14(12), 7 7. - Kosslyn, S. M., Ganis, G., & Thompson, W. L. (2001). Neural foundations of imagery. Nature Reviews Neuroscience, 2(9), 635. z, D. J., Saleem, K. S., Baker, C. I., Ungerleider, L. G., & Mishkin, M. (2013). The ventral Kravit visual pathway: an expanded neural framework for the processing of object quality. - Trends in cognitive sciences, 17(1), 26 49. - A. D. (2012). Multi Kuhl, B. A., Rissman, J., & Wagner, voxel patterns of visual category representation during episodic encoding are predictive of subsequent memory. 469. - Neuropsychologia, 50(4), 458 s Kuhl, B. A., Bainbridge, W. A., & Chun, M. M. (2012). Neural reactivation reveals mechanism for updating memory. Journal of Neuroscience, 32(10), 3453 - 3461. Lamme, V. A., Super, H., & Spekreijse, H. (1998). Feedforward, horizontal, and feedback - 535. processing in the visual cortex. Current opinion in neurobiology, 8(4), 529 LeCun, Y., Bengio, Y ., & Hinton, G. (2015). Deep learning. nature, 521(7553), 436. Lee, T. G., & D'Esposito, M. (2012). The dynamic nature of top - down signals originating from Journal of Neuroscience prefrontal cortex: a combined fMRI 32 (44), – TMS study. , 15458 15466. - Lee, S. dependent dissociation of visual and - H., Kravitz, D. J., & Baker, C. I. (2013). Goal prefrontal cortices during working memory. Nature neuroscience, 16(8), 997. Liu, L., Ouyang, W., Wang, X., Fieguth, P., Chen, J., Liu, X., & Pietikäinen, M. (2018). Deep

47 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . l earning for generic object detection: A survey. arXiv preprint arXiv:1809.02165. Luo, W., Li, Y., Urtasun, R., & Zemel, R. (2016). Understanding the effective receptive field in information processing deep convolutional neural networks. In Advances in neural 4906). - systems (pp. 4898 Mayer, J. S., Bittner, R. A., Nikolić, D., Bledowski, C., Goebel, R., & Linden, D. E. (2007). , Neuroimage (2), Common neural substrates for visual working memory and attention. 36 441 - 453. , Ma Negative Lasso and Elastic Net Penalized Generalized - J. (2016) . Non Mandal , B. N., & project.org/package=nnlasso - https://CRAN.R . URL: Linear Models . dependent Mante, V., Sussillo, D., Shenoy, K. V., & Newsome, W. T. (2013). Context - computation by recurrent dynamics in prefrontal cortex. nature, 503(7474), 78. - Mechelli, A., Price, C. J., Friston, K. J. & Ishai, A. (2004). Where bottom up meets top - down: . neuronal interactions during perception and imagery. Cereb C ortex , 14, 1256 ral – 65 Meyers, E. M., Freedman, D. J., Kreiman, G., Miller, E. K., & Poggio, T. (2008). Dynamic population coding of category information in inferior temporal and prefrontal cortex. Journal of neurophysiology, 100(3), 1407. Miller, E. K., Erickson, C. A., & Desimone, R. (1996). Neural mechanisms of visual working - 5167. memory in prefrontal cortex of the macaque. Journal of neuroscience, 16(16), 5154 Mumford, J. A., Turner, B. O., Ashby, F. G., & Poldrack, R. A. (2012). Deconvolving BOLD related designs for multivoxel pattern classification analyses. - ivation in event act

48 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . Neuroimage, 59(3), 2636 2643. - wise - Naselaris, T., Olman, C. A., Stansbury, D. E., Ugurbil, K., & Gallant, J. L. (2015). A voxel ental images of remembered scenes. encoding model for early visual areas decodes m - , Neuroimage 105 228. , 215 Nobre, A. C., Coull, J. T., Maquet, P., Frith, C. D., Vandenberghe, R., & Mesulam, M. M. (2004). Orienting attention to locations in perceptual versus mental representations. Journal of cognitive neuroscience, 16(3), 363 - 373. Pearson, J., Naselaris, T., Holmes, E. A., & Kosslyn, S. M. (2015). Mental imagery: functional - 602. mechanisms and clinical applications. Trends in cognitive sciences, 19(10), 590 down influences - Piëch, V., Li, W., Reeke, G. N., & Gilbert, C . D. (2013). Network model of top on local gain and contextual interactions in visual cortex. Proceedings of the National Academy of Sciences, 201317019. nference Poldrack, R. A. (2011). Inferring mental states from neuroimaging data: from reverse i 697. to large - scale decoding. Neuron, 72(5), 692 - Polyn, S. M., Natu, V. S., Cohen, J. D., Norman, K. A. (2005). Category - specific cortical activity precedes retrieval during memory search. Science. 310:1963 6. - ttention network of the human brain: action, saliency, and a Ptak, R. (2012). The frontoparietal a 515. - priority map of the environment. The Neuroscientist, 18(5), 502 Ptak, R., Schnider, A., & Fellrath, J. (2017). The dorsal frontoparietal network: a core system for emulated action. - 599. Trends in Cognitive Sciences, 21(8), 589

49 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . Rao, R. P., & Ballard, D. H. (1999). Predictive coding in the visual cortex: a functional field effects. Nature neuroscience, 2(1), - classical receptive - interpretation of some extra 79. - Raposo, D., Kaufman, M. T free neural population ., & Churchland, A. K. (2014). A category - making. Nature neuroscience, 17(12), 1784. supports evolving demands during decision Reddy, L., Tsuchiya, N., & Serre, T. (2010). Reading the mind's eye: decoding category information during 825. - mental imagery. Neuroimage, 50(2), 818 Riggall, A. C., & Postle, B. R. (2012). The relationship between working memory storage and elevated activity as measured with functional magnetic resonance imaging. Journal of 12998. Neuroscience, 32(38), 12990 - Rigotti, M., Barak, O., Warden, M. R., Wang, X. J., Daw, N. D., Miller, E. K., & Fusi, S. (2013). The importance of mixed selectivity in complex cognitive tasks. Nature, 497(7451), 585. Rissman, J., and Wagner, A. D. (2012). Distributed representations in memory: Insights from 28. functional brain imaging. Annual Review of Psychology . 63:101 – Rolls, E. T., Aggelopoulos, N. C., & Zheng, F. (2003). The receptive fields of inferior temporal cortex neurons in natural scenes. Journal of Neuroscience, 23(1), 339 - 3 48. selection and action selection - - Rowe, J., Hughes, L., Eckstein, D., & Owen, A. M. (2008). Rule have a shared neuroanatomical basis in the human prefrontal and parietal cortex. Cerebral 2285. - Cortex, 18(10), 2275 Saad, Z. S., Glen, D. R., Chen, G., Beauc hamp, M. S., Desai, R., & Cox, R. W. (2009). A new

50 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . method for improving functional to - structural MRI alignment using local Pearson - 848. - correlation. Neuroimage, 44(3), 839 . E., & van Seeliger, K., Fritsche, M., Güçlü, U., Schoenmakers, S., Schoffelen, J. M., Bosch, S - Gerven, M. A. J. (2018). Convolutional neural network based encoding and decoding of - visual object recognition in space and time. NeuroImage, 180, 253 266. - Shrivastava, A., Sukthankar, R., Malik, J., & Gupta, A. (2016). Beyond skip connections: Top down modulation for object detection. arXiv preprint arXiv:1612.06851. Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large - scale image recognition. arXiv preprint arXiv:1409.1556. Slotnick, S. D., Thompson, W. L., Kosslyn, S. M. (2005). Visual mental imagery induces 83. retinotopically organized activation of early visual areas. Cerebral Cortex. 15:1570 - Smith, A. T., Singh, K. D., Williams, A. L., & Greenlee, M. W. (2001). Estimating receptive uman striate and extrastriate visual cortex. Cerebral cortex, field size from fMRI data in h - 11(12), 1182 1190. Smith, F. W., & Goodale, M. A. (2013). Decoding visual object categories in early - 1031. somatosensory cortex. Cerebral cortex, 25(4), 1020 , Buchsbaum, B. R. (2014). Memory reactivation in healthy St - Laurent, M., Abdi, H., Bondad, A. aging: evidence of stimulus - specific dedifferentiation. Journal of Neuroscience. 34:4175 - 86. ) . , Buchsbaum , St . R. ( 2015 H . Distributed patterns of reactivation predict , , Abdi . M , Laurent - B

51 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . vividness of recollection. Journal of Cognitive Neuroscience. 27:2000 18 - Vandenberghe, R., Price, C., Wise, R., Josephs, O., & Frackowiak, R. S. J. (1996). Functional rds and pictures. Nature, 383(6597), 254. anatomy of a common semantic system for wo Blagoev, E. J., Clark, J., & Poldrack, R. A. (2001). Recovering meaning: - Wagner, A. D., Paré 338. left prefrontal cortex guides controlled semantic retrieval. Neuron, 31(2), 329 - Walther, D. B., Caddigan, E., Fei - Fe i, L., & Beck, D. M. (2009). Natural scene categories revealed in distributed patterns of activity in the human brain. Journal of neuroscience, 29(34), 10573 10581. - Warden, M. R., & Miller, E. K. (2010). Task dependent changes in short - - term memory in the 15810. - refrontal cortex. Journal of Neuroscience, 30(47), 15801 p Wen, H., Shi, J., Zhang, Y., Lu, K. H., Cao, J., & Liu, Z. (2017). Neural encoding and decoding - 25. with deep learning for dynamic natural vision. Cerebral Cortex, 1 Wynn, J. S., Bone, M. B., Dragan, M. C., Hoffman, K. L., Buchsbaum, B. R., Ryan, J. D. (2016). - Selective scanpath repetition during memory guided visual search. Visual Cognition. 24:15 - 37. , & DiCarlo, J. J. (2014). Yamins, D. L., Hong, H., Cadieu, C. F., Solomon, E. A., Seibert, D. - Performance optimized hierarchical models predict neural responses in higher visual 8624. - cortex. Proceedings of the National Academy of Sciences, 111(23), 8619 Zanto, T. P., Rubens, M. T., Bollinger, J., & Gazzaley, A. (2010). T - op down modulation of visual feature processing: the role of the inferior frontal junction. Neuroimage, 53(2),

52 bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license . 736 745. - Zanto, T. P., Rubens, M. T., Thangavel, A., & Gazzaley, A. (2011). Causal role of the prefrontal cortex in top - down modulation of visual processing and working memory. Nature neuroscience, 14(5), 656. Zhang, P., Wang, D., Lu, H., Wang, H., & Ruan, X. (2017, October). Amulet: Aggregating multi - level convolutional features for salient object detection. In Proceedings of the IEEE 211). - l Conference on Computer Vision (pp. 202 Internationa

Related documents

Feature Specific Neural Reactivation during Episodic Memory

Feature Specific Neural Reactivation during Episodic Memory

bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted b...

More info »
Feature Specific Neural Reactivation during Episodic Memory

Feature Specific Neural Reactivation during Episodic Memory

bioRxiv preprint first posted online Apr. 30, 2019; doi: http://dx.doi.org/10.1101/622837 . The copyright holder for this preprint (which was not peer-reviewed) is the author/funder, who has granted b...

More info »
CityNT2019TentRoll 1

CityNT2019TentRoll 1

STATE OF NEW YORK 2 0 1 9 T E N T A T I V E A S S E S S M E N T R O L L PAGE 1 VALUATION DATE-JUL 01, 2018 COUNTY - Niagara T A X A B L E SECTION OF THE ROLL - 1 CITY - North Tonawanda TAX MAP NUMBER ...

More info »
Out of Reach 2018

Out of Reach 2018

2018 of OUT REACH THE HIGH COST OF HOUSING MADE POSSIBLE BY THE GENEROSITY OF:

More info »
Out of Reach 2016

Out of Reach 2016

No Refuge for Low Income Renters MADE POSSIBLE BY THE GENEROSITY OF:

More info »
DER Directory

DER Directory

FAA CONSULTANT DER DIRECTORY May 9, 2019 AIR-6F0, Delegation & Organizational Procedures Branch This directory is generated from information in the FAA Designee Information Network (DIN). If you are a...

More info »
time is the enemy

time is the enemy

is TIME the ENEMY The surprising truth about why today’s college students graduating ... and wha T needs To change aren’t n 1 Time Is the Enemy

More info »
50 Year Trends expanded version

50 Year Trends expanded version

Expanded Edition THE AMERICAN HERI A R CIRP T B I E N L G E C 50 F S I F R T A Y E Y FRESHMAN FIFTY-YEAR TRENDS | 1966-2015 KEVIN EAGAN | ELLEN BARA STOLZENBERG | JOSEPH J. RAMIREZ | MELISSA C. ARAGON...

More info »
Implementation Handbook For The Convention On The Rights Of The Child

Implementation Handbook For The Convention On The Rights Of The Child

IMPLEMENTATION HANDBOOK FOR THE CONVENTION ON THE RIGHTS OF THE CHILD FULLY REVISED THIRD EDITION IMPLEMENTATION HANDBOOK IMPLEMENTATION HANDBOOK FOR THE CONVENTION ON THE FOR THE CONVENTION ON THE RI...

More info »
Abstract and Concrete Categories   The Joy of Cats

Abstract and Concrete Categories The Joy of Cats

Jiˇr ́ı Ad ́amek Horst Herrlich George E. Strecker Abstract and Concrete Categories The Joy of Cats Dedicated to Bernhard Banaschewski

More info »
IWMF Global Report

IWMF Global Report

Global Report on the Status of Women in the News Media INTERNATIONAL WOMEN’S MEDIA FOUNDATION

More info »
What's It Worth? The Economic Value of College Majors

What's It Worth? The Economic Value of College Majors

Anthony P. Carnevale Jeff Strohl Michelle Melton

More info »
S:\FULLCO~1\HEARIN~1\Committee print 2018\Henry\Jan. 9 report

S:\FULLCO~1\HEARIN~1\Committee print 2018\Henry\Jan. 9 report

1 C . RT S. P 115 TH ONGRESS " ! COMMITTEE PRINT 2d Session 115–21 PUTIN’S ASYMMETRIC ASSAULT ON DEMOCRACY IN RUSSIA AND EUROPE: IMPLICATIONS FOR U.S. NATIONAL SECURITY A MINORITY STAFF REPORT PREPARE...

More info »
AndersBehringBreivikManifesto

AndersBehringBreivikManifesto

2011 , London – By Andrew Berwick

More info »
DoD7045.7H

DoD7045.7H

DoD 7045.7-H EPARTMENT OF D EFENSE D F UTURE Y EARS D EFENSE P ROGRAM (FYDP) S TRUCTURE Codes and Definitions for All DoD Components Office of the Director, Program Analysis and Evaluation A pril 2004

More info »
Web Tables—Profile of Undergraduate Students: 2011–12

Web Tables—Profile of Undergraduate Students: 2011–12

Profile of WEB Undergraduate 12 Students: 2011– TABLES EDUCATION OF DEPARTMENT U.S. NCES 2015-1 OCTOBER 2014 67 These Web Tables provide comprehen- distance from home; and participation carried a bala...

More info »
Microsoft Word   cust subm Book 251 Cover.doc

Microsoft Word cust subm Book 251 Cover.doc

Summary of Reciprocity Agreements Between Texas and Other Jurisdictions Vehicle Titles and Registration Division Austin, Texas Issued November 2002

More info »
summer schedule full

summer schedule full

Summer Semester 2019 Class Schedule May 9, 2019 as of Subj Crse CRN Course Type Title Days Time Bldg/Rm Cred Instructor Course Fee Start Date - End Date Campus Part Term BB 321 Financial Accounting I ...

More info »
An Introduction to Computer Networks

An Introduction to Computer Networks

An Introduction to Computer Networks Release 1.9.18 Peter L Dordal Mar 31, 2019

More info »