Priority-based transformations of stimulus representation in visual working memory.

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience

Background:

• Understanding how the brain prioritizes information in working memory (WM) is crucial for explaining goal-directed behavior.
• Prior research using EEG and fMRI suggests unprioritized memory items (UMIs) are represented in an inverted format compared to prioritized memory items (PMIs).

Purpose of the Study:

• To provide independent evidence for priority-based representational transformations in WM.
• To explore the underlying neural mechanisms of this representational transformation using computational models.

Main Methods:

• Recurrent neural networks (RNNs) with a Long Short-Term Memory (LSTM) architecture were trained on a 2-back WM task.
• Principal Component Analysis (PCA) and Demixed PCA (dPCA) were used to visualize and analyze hidden layer activity in the LSTMs.
• dPCA was also applied to EEG data to compare findings with neural data.

Main Results:

• LSTM hidden layer activity visualized with PCA showed stimulus representations transformed, consistent with a flip when transitioning from UMI to PMI.
• dPCA revealed distinct representational trajectories within UMI and PMI subspaces, both involving a reversal of stimulus coding axes.
• EEG data analyzed with dPCA also supported priority-based transformations, though with some variations.

Conclusions:

• The findings suggest a representational transformation mechanism underlies WM prioritization, allowing retention of UMIs without behavioral interference.
• RNN models provide a valuable tool for investigating WM mechanisms, though algorithmic differences between models and the brain warrant further study.