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.
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.
Fig 3.
Example neural activations and their corresponding decoder curves.
Top row: neural activity corresponding to a single trial condition. The leftmost panel is the actual neural data. The other panels are artificial neural networks, whose details are described in the main text. Bottom row: decoder accuracy curves corresponding to the neural activations in the top row.
Fig 4.
a) A decoder is trained to predict the sample label from RNN neural trajectories, for the fixed-synapse RNNs. a1) The accuracy of the trained decoder as a function of time from sample onset for FS-tanh. a2) The same plot as (a1), for FS-relu. b) Two decoders are trained on the neural and synaptic trajectories separately. Black lines indicate neural decoding, purple indicate synaptic decoding. b1) Neural and synaptic decoding accuracy as a function of time for PS-pre. b2) Same plot as in (b1) but for PS-hebb.
Fig 5.
Results of the hyperparameter sweep across number of hidden neurons, parameter regularization strength, and activity regularization strength.
Each row corresponds to a different parameter regularization (1e-4,1e-3,1e-2). Each column corresponds to a different observed quantity (brain-likeness, structural robustness, process robustness). Each subplot is a 10x10 grid, corresponding to 10 possible hidden size / activity regularization configurations. Shown in each square of that grid is the most brain-like/robust network corresponding to that particular hyperparameter configuration. LSTM and GRU networks were excluded (see S8 Fig for corresponding figure when they are included). Each color corresponds to a different network.
Fig 6.
Dimensionality reduced space plots for the most brain-like models.
Time-averaged RNN activity during the sample-period as well as 500ms before the end of the delay period. Linear Discriminant Analysis (LDA) was used to project the data into three-dimensions. To account for differences in training/test splits, an Orthogonal Procrustes operation was used to rotationally align the sample and delay period activity. Colors denote sample IDs. Fixed synapse models have activity organized around simple attractors in state space. There is one attractor for each sample ID. Plastic synapse models exhibit high sample-separability in synaptic state space, and limited separability in the neural state space during the delay period. Similarly, PFC exhibited higher neural discriminability during the sample than during the delay period.
Fig 7.
Distances between neural trajectories within a sample condition and between sample conditions, for fixed synapse RNNs.
All models used were the most ‘brain-like’, as determined by the methodology in section “RNNs With STSP are More Brain-Like”. a) Cartoon of trial-averaged RNN trajectories corresponding to two different sample conditions for the fixed-synapse RNNs. a1) Average pairwise distance between trajectories on different sample conditions, for the fixed synapse network with tanh activation (FS-tanh). a2) The same plot as in a1, but for FS-relu. b1) The average distance between distracted and undistracted trajectories. Average taken over all sample conditions. Results shown for FS-tanh. b2) Same results as in b1, but for FS-relu.
Fig 8.
Distances between neural and synaptic trajectories within a sample condition and between sample conditions, for plastic synapse RNNs.
Black lines correspond to neural trajectories, purple lines correspond to synaptic trajectories. a) Cartoon of trial-averaged neural and synaptic RNN trajectories corresponding to two different sample conditions for the plastic-synapse RNNs. a1) Average pairwise distance between neural and synaptic trajectories on different sample conditions, for PS-pre. a2) The same plot as in a1, but for PS-hebb. b1) The average distance between distracted and undistracted trajectories. Average taken over all sample conditions. Results shown for PS-pre. b2) Same results as in b1, but for PS-hebb.