Mathematical Models from AI Research Help Explain How Our Brain Remembers

Our memory is far from infallible – in fact, we sometimes recall things that never actually happened. From the perspective of memory research, one advantage of memory errors is that they tend to occur in systematic ways. This regularity offers a unique opportunity to uncover the underlying mathematical principles. These principles also hold the promise of helping us understand how various optimisation processes in memory can lead to errors.

In a study published in Nature Reviews Psychology, Gergő Orbán (HUN-REN Wigner Research Centre for Physics), in collaboration with Dávid Gergely Nagy and Charley Wu from the Max Planck Institute and the University of Tübingen, proposes a machine learning-based framework for understanding the interactions between human memory systems. But how is machine learning relevant here? Mathematical models originating from artificial intelligence research provide tools not only to document memory errors, but also to reveal their functional role – offering insights into the learning and information compression principles that give rise to them.

Information theory offers guidance on what is worth remembering – that is, which pieces of information are worth the cost of storage, and which are better discarded. In their study, the authors point out that, according to information-theoretic principles, it would not be ‘optimal’ for the brain to retain extremely rare or atypical experiences. Yet these surprising events often leave particularly deep and vivid traces in memory.

Based on mathematical considerations, the authors conclude that in order to support effective learning, the brain must retain surprising and unusual experiences. While not entirely extraordinary, these events are distinctive enough to stand out from the everyday – and it is precisely this distinctiveness that helps us build a better understanding of how the world works.

While storing past experiences generally helps us plan and predict more effectively, the memorisation of surprising events plays a critical role in updating our knowledge – ensuring that we can continue to plan efficiently in the future. Machine learning not only helps predict what our memory is likely to retain or forget, but also provides guidance on how to learn or teach more effectively: when to revisit and when to move on to the next challenge.