(C) PLOS One This story was originally published by PLOS One and is unaltered. . . . . . . . . . . Robust and brain-like working memory through short-term synaptic plasticity [1] ['Leo Kozachkov', 'The Picower Institute For Learning', 'Memory', 'Massachusetts Institute Of Technology', 'Mit', 'Cambridge', 'Massachusetts', 'United States Of America', 'Department Of Brain', 'Cognitive Sciences'] Date: 2023-01 Working memory has long been thought to arise from sustained spiking/attractor dynamics. However, recent work has suggested that short-term synaptic plasticity (STSP) may help maintain attractor states over gaps in time with little or no spiking. To determine if STSP endows additional functional advantages, we trained artificial recurrent neural networks (RNNs) with and without STSP to perform an object working memory task. We found that RNNs with and without STSP were able to maintain memories despite distractors presented in the middle of the memory delay. However, RNNs with STSP showed activity that was similar to that seen in the cortex of a non-human primate (NHP) performing the same task. By contrast, RNNs without STSP showed activity that was less brain-like. Further, RNNs with STSP were more robust to network degradation than RNNs without STSP. These results show that STSP can not only help maintain working memories, it also makes neural networks more robust and brain-like. Working memory has been thought to depend on sustained spiking alone. But recent evidence shows that spiking is often sparse, not sustained. Short-term synaptic plasticity (STSP) could help by maintaining memories between spiking. To test this, we compared artificial recurrent neural networks (RNNs) with and without short-term synaptic plasticity (STSP). Both types of RNNs could maintain working memories. But RNNs with STSP functioned better. They were more robust to network degradation. Plus, their activity was more brain-like than RNNs without STSP. These results support a role for STSP in working memory. Funding: This work was support by Office of Naval Research N00014-22-1-2453 (E.K.M), The JPB Foundation (EK.M.), ERC Starting Grant 949131 (M.L), and VR Starting Grant 2018-04197 (M.L). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Copyright: © 2022 Kozachkov et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. We trained Recurrent Neural Networks (RNNs) with and without STSP to test how it affects network performance and function. We focused on the key property of robustness [ 9 , 24 – 27 ]. Working memories must be maintained in the face of distractions. Networks need to deal with noise and show graceful degradation (i.e., continue to function when portions of the network are damaged). Our analysis showed RNNs with and without STSP were robust against distractors. However, only the RNNs with STSP were “brain-like”—their activity more closely resembled activity recorded from the prefrontal cortex of a NHP performing a WM task. RNNs with STSP were also more robust against synaptic ablation. Thus, STSP offers functional advantages and explains how WM can be maintained between stimulus presentations. This point has been made repeatedly in the literature. However, our study is the first to train artificial neural networks with STSP and quantitatively measure their similarity to recorded electrophysiological data. This all begs the question of how WMs are maintained over these gaps in time with little-to-no spiking [ 11 , 16 ]. One possibility was suggested by observations of short-term synaptic plasticity (STSP), transient (< 1 second) changes in synaptic weights induced by spiking, in circuits in the prefrontal cortex [ 17 ]. Several groups have suggested updating the attractor dynamics model with this feature [ 18 – 20 ]. The idea is that STSP helps the spiking. Spikes induce a transient “impression” in the synaptic weights that can maintain the network state between spikes [ 21 – 23 ]. Evidence for STSP comes from techniques like patch-clamp recording that are difficult to implement in the working brain, especially in NHPs. Thus, we tested the role of STSP in WM by using computational modeling in conjunction with “ground-truthing” via analysis of spiking recorded from the PFC of a NHP performing a WM task. We found that PFC spiking carried little-to-no stimulus-specific WM information across the delay. We aimed to determine if network models with STSP can solve the working memory task, whether they have properties similar to those seen in the actual brain, and whether STSP endows functional advantages. The answer to these questions was “yes”. Working memory (WM), the holding of information “online” and available for processing, is central to higher cognitive functions [ 1 , 2 ]. A well-established neural correlate of WM is spiking over a memory delay [ 3 – 5 ]. For many years, this was thought to be the sole mechanism underlying WM maintenance. The idea is that sensory inputs elicit unique patterns of spiking that are sustained via recurrent connections [ 6 ], creating attractor states—stable patterns of activity that retain the WM [ 7 ]. It seems evident that these attractor dynamics play an important role in WM. Recent observations, however, have suggested that there may be more going on [ 8 – 11 ]. A few neurons seem to show spiking that looks persistent enough to be an attractor state, but the bulk of neurons show memory delay spiking that is sparse [ 12 – 15 ]. This is especially true when spiking is examined in real time (i.e., on single trials) because averaging across trials can create the appearance of persistence even when the underlying activity is quite sparse. Results A NHP was trained to perform an object delayed-match-to-sample task (Fig 1). The NHP was shown a sample object and had to choose its match after a variable-length memory delay. At mid-delay a distractor object (1 of 2 possible objects never used as samples) was presented (for 0.25s) on 50% randomly chosen trials. We recorded multi-unit activity (MUA) bilaterally in dorsolateral PFC (dlPFC) and ventrolateral PFC (vlPFC) using four 64-electrode Utah arrays for a total of 256 electrodes. The animal learned to do the task consistently at ~99% accuracy for both distractor and non-distractor trials. PPT PowerPoint slide PNG larger image TIFF original image Download: Fig 1. Electrode location and task structure. Utah arrays were implanted bilaterally in dorsolateral PFC (dlPFC) and ventrolateral PFC (vlPFC). Animal performed a distracted delayed match-to-sample task. Each trial began with visual fixation on the middle of the screen for 0.5s. Fixation was maintained throughout the trial until the behavioral response. The delay length was parametrically varied from 1–4 s in five logarithmic steps, randomly chosen each trial. At mid-delay a neutral distractor (1 of 2 possible objects never used as samples) was presented randomly on 50% of trials. During the multi-choice test the NHP was allowed to freely saccade between all objects on the screen. The final choice was indicated by fixating on it for at least one second. https://doi.org/10.1371/journal.pcbi.1010776.g001 Sample information in population neural activity was weak over longer delays First, we examined MUA recorded from the lateral PFC. To quantify the amount of sample object information carried by spiking, we used a linear classifier (see Methods for details)1. This showed that spiking carried sample object information for about one second after the sample disappeared. From the start of the delay period the decoder accuracy decreased steadily towards chance (Figs 2A and S1). We corroborated this by measuring the distance between neural population activity for all pairs of sample objects. That gave an average distance between experimental conditions at every timepoint (see Methods). This showed that the distance between population MUA activity for different samples returned to pre-stimulus levels (Figs 2B and S2). Interestingly, we found that this was not simply due to spiking returning to pre-sample values. We determined this by training a classifier to discriminate between pre and post sample spiking activity. We found that this classifier was consistently able to discriminate between pre and post sample spiking activity over the delay (S3 Fig). PPT PowerPoint slide PNG larger image TIFF original image Download: Fig 2. a) Left: training a decoder to predict sample identity given a neural trajectory. Right: decoder accuracy on held-out trials for distracted vs. undistracted trials b) Left: comparing trial-averaged trajectories corresponding to different samples. Right: average pairwise distance in state-space between trajectories elicited by all possible sample images. Normalized by the average pre-stim distance. c) Left: comparing trial-averaged distracted vs. non-distracted trajectories through neural state space. Right: distance in state space between distracted vs. non-distracted trajectories throughout the trial. Shown are trials with a delay of four seconds. https://doi.org/10.1371/journal.pcbi.1010776.g002 [END] --- [1] Url: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1010776 Published and (C) by PLOS One Content appears here under this condition or license: Creative Commons - Attribution BY 4.0. via Magical.Fish Gopher News Feeds: gopher://magical.fish/1/feeds/news/plosone/