The Hot-Coal hypothesis of working memory

Working memory (WM) is a fundamental cognitive capability. It refers to our ability to hold, select and manipulate several objects in mind simultaneously. It allows us to engage in flexible behavior and is tightly linked to fluid intelligence.

This project will answer an essential, yet unsolved aspect of WM: How can primates use their WM in a generalized way and control what they think about?

If you hear apple, stone and pear in sequence, and then you are asked to imagine the first fruit, how is it that you do not confuse apples with pears? There are many competing models of WM, but no biologically detailed models are capable of generalization.

Neural networks can be trained to perform similar WM tasks as primates do, a major difference is that primates generalize their training. They can learn the task on a set of objects, then perform it on a novel set.

Computational models typically rely on changing the connections between units to achieve the desired activity patterns to solve the task. Since these activity patterns depend on the objects held in WM, the training does not translate to novel objects. I propose a new solution to this problem, the Hot-Coal model of WM.

It relies on a novel computational principle in which spatial location of information, rather than connectivity, is controlled by excitatory bursts to support cognition. I will explore this principle and test it in data. Preliminary tests suggest that the Hot-Coal theory is supported by electrophysiological data from primates.

By implementing the theory in computational networks I aim to demonstrate the generalization mechanism and provide more detailed predictions.

Finally, I will use the theory to resolve seemingly conflicting findings regarding the mechanisms underlying WM, by reproducing them in a single model.

The new theory could constitute a significant advance in the mechanistic understanding of one of the most central and puzzling components of cognition.