Adaptive chunking improves effective working memory capacity in a prefrontal cortex and basal ganglia circuit
- Brown University, United States
Figures
Figure 1
Base model sensory inputs (reflecting visual cortical representations) project to PFC superficial layers, which transiently represent those inputs.
Activity is maintained in prefrontal cortex (PFC) after stimulus offset (and into subsequent trials) only when gated. Red arrows indicate gating, supporting transfer of information from superficial …
Figure 2
Visual working memory task.
(a) The color wheel task is commonly used to study the nature of capacity limitations in VWM. During encoding, participants are presented multiple randomly generated oriented and colored bars. After …
Figure 3
Example sequence of network gating decisions.
In this example trial, the network is presented with stimulus 1 (color and orientation), stimulus 2, and is then asked to recall the color of stimulus 1 based on just its orientation. Each step is …
Figure 4
Prefrontal cortex basal ganglia working memory (PBWM) model and chunking layer details.
(a) Network diagram in the minimal case of two stripes: the first prefrontal cortex (PFC) stripe receives projections from the input layer (‘PFC Input Stripe’); the second PFC stripe receives …
Figure 5 with 1 supplement
Model recall error histograms.
The binned error in degrees is plotted on the x-axis, and the number of trials for that error bin on the y-axis. The blue and orange histograms show errors from all recall trials across all 80 …
Figure 5—figure supplement 1
P(Recall) across set size.
(a) Average recall probability across set sizes decreases with set size, but less so for chunk models. Note that chance performance is approximately 19%. (b) Chunk models have a higher ratio of …
Figure 6
Chunking improves recall for non-chunked items.
Left. Example array. Here, we compare two sets, both containing a red item that will be later probed. In Set A, the other items (out of the probed cluster) are two shades of green and thus low …
Figure 7
Stripe usage.
(a) Stripe usage for the (1) chunk model, chunk -linked stripe (2) chunk model, input-linked stripe (3) no chunk model (average across both stripes), (b) Proportion of trials when at least one …
Figure 8
Increasing allocated capacity ≠ better performance.
The importance of Resource Management (a) Chunk model with four stripes vs. No chunk model with four stripes in a task with set size 4. Even though the no-chunk networks has sufficient number of …
Figure 9
Gating policy (Go - NoGo Weights for Each PFC Stripe) across training as the networks learn (over 500 training epochs, averaged over 80 networks), the learned gating strategy differentiates between the input-linked (orange) or chunk-linked (blue) stripes.
Positive values indicate the networks learn greater Go than NoGo weights for input gating stimuli into the corresponding stripe. (a) Set size 2, the learned gating strategy shows a slight preference …
Figure 10
Dynamic dopamine bursts and dips are needed for adaptive performance.
Each box is an average absolute error over 80 models. The color bar on the right indicates performance (note different scales on each plot), with darker colors (blue) representing better …
Figure 11
Network captures recency effects.
Average error on recall trials as a function of the distance in trials between presentation of the relevant stimulus and recall.
