What is it about?

Adapting to a complex and rapidly changing environment requires maintaining and constantly updating our working memory (WM). While neural mechanisms underlying WM maintenance have been extensively studied, how the brain achieves efficient updating has remained elusive. A key problem lies in the conflicting demands involved in maintenance and updating: While maintenance requires stability against noise and distractors, updating (incorporating new information) requires sensitivity to environmental changes (i.e., instability). This paper analyzes this problem from a dynamical systems perspective. Using monkeys performing a group reversal task and a trained recurrent neural network model, we demonstrate that stability and sensitivity can coexist in a single dynamic process unfolding at the population level, providing a mechanism for updating.

Featured Image

Why is it important?

This mechanism reveals a new type of neural representation, a “branching channel”, in the system’s state space. By “concatenating” such channels, a neural system might be able to build complex hierarchical structures (e.g., a decision tree). Together, these results provide a framework for understanding higher cognitive functions and suggest a reconciliation between the symbolic and connectionist approaches for building intelligent machines.

Perspectives

Two main paradigms have emerged in the history of pursuing human intelligence. The connectionist paradigm, which uses neural network models, has recently achieved great success, but its ability to perform higher cognitive functions, such as reasoning, is still limited. The symbolic paradigm, which postulates propositional representations for reasoning, suffers from the problem of learning such representations from data. Here, by uncovering the neural mechanism underlying WM updating, we provide insights into how these two paradigms might be reconciled. We hope that our study facilitates future efforts in this direction and helps integrate analyses across different levels (e.g., Marr’s three levels) of description.

Muyuan Xu
International Research Center for Neurointelligence, The University of Tokyo

Read the Original

This page is a summary of: Dynamic tuning of neural stability for cognitive control, Proceedings of the National Academy of Sciences, November 2024, Proceedings of the National Academy of Sciences,
DOI: 10.1073/pnas.2409487121.
You can read the full text:

Read

Contributors

The following have contributed to this page