Dr Rui Ponte Costa, Lecturer in Neuroscience at LMH and Associate Professor in the University of Oxford’s Department of Physiology, Anatomy & Genetics (DPAG), is the corresponding author of a new study published in Nature Communications that examines how the hippocampus supports flexible learning when sensory information is uncertain or incomplete.
The research, titled ‘Hippocampus supports multi-task reinforcement learning under partial observability’, was led by Dr Ponte Costa, DPAG’s Dr Dabal Pedamonti and Dr Samia Mohinta (Cambridge). They collaborated with researchers at the University of Vienna (Hugo Malagon-Vina) and the University of Bern (Stephane Ciocchi). Their article reports findings from behavioural experiments in rodents, computational modelling and neural recordings, through which the team aimed to uncover how hippocampal circuits enable learning in complex environments.
Their results show that when navigating environments in which key cues are hidden or ambiguous, animals switch between egocentric (movement-based) and allocentric (environment-based) spatial strategies. Deep learning models inspired by hippocampal circuitry demonstrated that recurrent, memory-based connections were essential for learning under ‘partial observability’ conditions. Models lacking recurrence failed to adapt when information was missing or unclear.
Recordings from hippocampal neurons closely matched the internal dynamics of the computational models, with both reflecting information related to strategy, timing and decision-making. This suggests that hippocampal recurrence plays a key role in inferring hidden aspects of the environment and guiding behavioural choices.
The study highlights the importance of examining brain mechanisms in complex, partially observable settings that more closely resemble real-world conditions, and it contributes to ongoing work at the interface of neuroscience and artificial intelligence, pointing to new ways of designing AI systems that can adapt flexibly in real-world settings. The paper was recently highlighted by Nature Neuroscience.
Dr Ponte Costa leads the University of Oxford’s Neural & Machine Learning group, which develops computational models of learning in the brain, focusing on cortical circuits, neuromodulation and subcortical processes involved in credit assignment.
