How Generative AI Illuminates Memory, Imagination, and Planning — UCL Study

Summary: A new study from UCL uses generative AI to model how the human brain encodes, stores, and reconstructs memories for learning, imagination, and planning. The researchers built a computational network that mirrors interactions between the hippocampus and neocortex to simulate how the brain captures individual experiences and extracts broader conceptual knowledge.

Using a generative neural network, the study shows how the neocortex can form compact, conceptual representations from repeated experience and how the hippocampus supports rapid encoding of individual events. The model demonstrates how these complementary processes allow the brain both to re-create past events with distinctive details and to generate novel scenarios useful for predicting future outcomes.

Key Facts:

1. The computational model simulates the interaction between hippocampal and neocortical networks during memory encoding, replay, and retrieval.
2. The neocortex learns efficient “conceptual” representations that capture the meaning and structure of experiences, supporting reconstruction and creative generation.
3. The findings clarify how memory supports survival-related prediction and why memories often exhibit “gist-like” distortions when unique details are generalized.

Source: UCL

Recent progress in generative AI provided a framework for researchers at UCL to investigate how memories enable us to learn about the world, re-live past events, and construct entirely new experiences for imagination and planning. The study, published in Nature Human Behaviour and funded by Wellcome, applied a generative neural network to simulate the brain’s learning from a series of simple scenes.

The model included separate network components representing the hippocampus and the neocortex to study their interaction. Both brain regions are known to cooperate during memory formation, offline replay, and the flexible recombination of information needed for imagining or planning future scenarios.

Lead author Eleanor Spens (PhD student, UCL Institute of Cognitive Neuroscience) notes that modern generative networks illustrate how the brain can distill information from experience so we can both recall a specific episode and flexibly imagine novel possibilities. The research frames remembering as a form of imagining the past: combining some preserved details with conceptual expectations about what typically occurs in similar situations.

In the simulations, the researchers presented 10,000 simple scene images to the model. The hippocampal component rapidly encoded each scene as it was encountered, then replayed those encoded patterns repeatedly to train the generative network in the neocortex. Through replay, the neocortex learned to map thousands of input neurons (which convey sensory details) through much smaller hidden layers to thousands of output neurons that reconstruct or predict sensory patterns.

This compression into smaller intermediate layers encouraged the neocortex to form highly efficient, conceptual representations that capture scene structure and meaning—such as the arrangement of walls and objects—rather than preserving every low-level detail. With these compressed concepts, the neocortical network could both recreate previously seen scenes and generate entirely new scene variants.

As a result, the hippocampus could focus its rapid encoding on unique features of new experiences that the neocortex’s conceptual model could not reproduce—novel objects or unusual configurations—while relying on neocortical knowledge for more common, predictable aspects. This division of labor explains how the neocortex gradually acquires generalized knowledge, while the hippocampus supports specific, episodic encoding.

The model also clarifies why memories are often biased toward the “gist” of events. When reconstructing an episode, the brain recombines conceptual foundations with stored unique details; if those unique details are weak or missing, reconstructions will lean on the neocortex’s generalized patterns, producing systematic distortions that make memories seem more like prior experiences.

Senior author Professor Neil Burgess (UCL Institute of Cognitive Neuroscience and UCL Queen Square Institute of Neurology) emphasizes that these reconstructed memories are not faithful recordings but adaptive combinations of meaning and detail. This recombination helps explain how memory supports prediction, decision-making, and creative imagining, while also producing predictable biases in recall.

About this AI, imagination, and memory research news

Author: Poppy Danby
Source: UCL
Contact: Poppy Danby – UCL
Image: Image credited to Neuroscience News

Original Research: The findings will appear in Nature Human Behaviour