A recurrent neural network model of prefrontal brain activity during a working memory task

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience

Background:

• Retro-cues enhance short-term memory recall by prioritizing specific items.
• The neural mechanisms and computational principles underlying retro-cue effects remain unclear.
• Previous research observed representational geometry changes in macaque lateral prefrontal cortex (LPFC) after retro-cues.

Purpose of the Study:

• To investigate how representational geometry transformations, observed in LPFC during retro-cueing, can be learned through error minimization.
• To computationally model the effects of retro-cues on short-term memory using recurrent neural networks (RNNs).
• To explore the functional significance of representational geometry changes for memory maintenance and recall.

Main Methods:

• Trained supervised RNNs to perform a cued-recall task with conjunctive stimuli and retro-cues.
• Analyzed neural population activity geometry (orthogonal vs. parallel subspaces) in RNNs before and after retro-cue presentation.
• Investigated learning dynamics, connectivity patterns, and effects of probabilistic cues in the trained RNNs.

Main Results:

• The orthogonal-to-parallel geometry transformation observed in macaque LPFC spontaneously emerged in the trained RNNs.
• Parallel geometry, indicative of a common readout, developed specifically when information required extended maintenance, suggesting robustness.
• RNNs exhibited learning dynamics and connectivity patterns consistent with error-driven learning of representational transformations.

Conclusions:

• Recurrent neural networks can learn to implement the representational geometry changes observed in the brain during retro-cueing tasks.
• The emergence of parallel geometry in RNNs suggests it may serve to stabilize memory representations during extended maintenance.
• Findings support theoretical models proposing that retro-cues facilitate recall by transforming memory representations into an action-oriented format.