P152 Differing Strategies and Neural Representations in the Same Long-Term Information Encoding Task
Tomoki Kurikawa*1
1Department of Complex and Intelligence Systems, Future University Hakodate, Hakodate, Japan *Email: kurikawa@fun.ac.jp
Introduction
Many cognitive tasks require maintaining information across trials, such as deterministic or probabilistic reversal learning tasks. In the deterministic reversal learning task [1], for instance, the pairing between sensory cues and behavioral outcomes reverses after a fixed number of trials. To perform such a task successfully, subjects have to track the number of elapsed trials to predict reversals accurately. However, the neural representation underlying such sustained memory processes remains poorly understood.
Methods To uncover the representations underlying task performance, we built a simple recurrent neural network (RNN) model trained on a deterministic reversal learning task using machine learning techniques. We analyzed what representations emerged and how they were formed. In this task, there were two types of blocks, and depending on the block type, the network had to alternate between two outputs (Left and Right outputs). Each block consisted of 10 trials, and the block type switched every 10 blocks iteratively. Notably, no explicit contextual cues were provided—the network had to track trial counts internally. The model was trained to produce correct outputs across 10 consecutive blocks.
Results We found that two distinct strategies emerged after learning 10 blocks: generalization and specification. In the generalization strategy, the network discovered the underlying rule of the task. Despite being trained on only 10 blocks, it could generalize and perform correctly beyond this limit. In contrast, in the specification strategy, the network was specifically trained to complete the 10-block task but was unable to extend its performance to a larger number of blocks, such as a 20-block task. What representations underlie these different behaviors? Our analysis revealed that different neural representations support these distinct strategies. In the generalization strategy, certain neurons specifically encoded the number of trials within a block. Their activity gradually increased across trials, and when a threshold was reached, the network switched outputs from one output to another before resetting, indicating that these neurons tracked the number of trials within a block. In contrast, in the specification strategy, no individual neurons encoded trial counts explicitly. Instead, this information was distributed across the neural population, implying a different mechanism for task execution.
Discussion Our findings suggest that even when performing the same task, different strategies can emerge across subjects or animals. Depending on the adopted strategy, the way long-term information is encoded across trials also varies. This computational result provides new insights into how long-term information is represented in neural systems.
Acknowledgements The present work is supported bySpecialResearch Expenses in Future University Hakodate References https://doi.org/10.1523/ENEURO.0172-24.2024
