In a fresh study by researchers at UCL, the latest breakthroughs in generative AI shed light on the mechanisms behind our ability to learn, recollect past experiences, and forge entirely new imaginative scenarios. Published in Nature Human Behaviour, the study employs a generative neural network, a form of AI computational model, to mimic how the brain's neural networks assimilate and retain a sequence of events, each represented by a simple scene.
The model incorporates networks representing the hippocampus and neocortex, aiming to understand their collaboration, known to occur during memory, imagination, and planning. Lead author Eleanor Spens, a Ph.D. student at UCL's Institute of Cognitive Neuroscience, notes, "Recent advances in generative networks used in AI show how information can be extracted from experience, allowing us to both recollect specific events and imaginatively envision new ones."
Human survival necessitates making predictions, and the AI networks propose that replaying memories during rest assists the brain in recognizing patterns from past experiences, aiding in such predictions. The researchers exposed the model to 10,000 images of simple scenes, observing the rapid encoding of each scene by the hippocampal network. These scenes were then replayed to train the generative neural network in the neocortex.
The neocortical network learned to transmit activity from input neurons representing each scene through smaller layers of neurons, eventually generating patterns of activity in output neurons. This process allowed the neocortical network to develop efficient conceptual representations of scenes, enabling both the recreation of old scenes and the generation of entirely new ones.
Consequently, the hippocampus could encode the meaning of new scenes without detailing every single aspect, directing resources towards encoding unique features that the neocortex couldn't replicate, like new types of objects. This model elucidates how the neocortex gradually acquires conceptual knowledge, working in tandem with the hippocampus to enable us to "re-experience" events by reconstructing them mentally.
The model also clarifies how imagination and future planning involve generating new events and why existing memories often include distortions with generalized features, termed "gist-like" distortions. Professor Neil Burgess, the senior author from UCL's Institute of Cognitive Neuroscience and Queen Square Institute of Neurology, explains, "Memories are not veridical records of the past; instead, they are reconstructed, highlighting how the meaning of an experience combines with unique details, leading to biases in our recollections."
nature biomedical engineering
ai research