ANC Seminar - Rui Ponte Costa

Tuesday, 23rd January 2024

Unlocking efficient credit assignment in the brain

Abstract: 

The brain assigns credit to billions of synapses remarkably well. How the brain achieves this feat is one of the unsolved mysteries in neuroscience. Recently, we have developed a model of hierarchical credit assignment that captures a large number of biological features while approximating deep learning algorithms. Our model relies on long- and short-term synaptic plasticity, dendritic computations and excitatory-inhibitory cell-types to enable error propagation. I will show that in contrast to previous work this model (i) does not require a multi-phase learning process, (ii) is consistent with experimental observations and (iii) provides efficient credit assignment across multiple cortical areas.

However, credit assignment depends critically on behavioural feedback, which may not be always readily available. How the brain learns efficiently despite the sparse nature of feedback remains unclear. Inspired by recent deep learning algorithms we propose that a specialised brain region, the cerebellum, predicts cortical feedback, thereby unlocking learning in cortical networks from future feedback. When trained in a range of motor and cognitive tasks the model shows faster learning and reduced dysmetria-like behaviours, in line with the widely observed functional role of the cerebellum.

Overall, our work suggests that synaptic, sub-cellular, cellular and systems features jointly underlie efficient credit assignment in the brain.

Bio: Rui leads the Neural & Machine Learning group at the University of Oxford. His group builds  AI-driven models of learning to transform our understanding of how the brain as a whole learns to adapt to the environment. The group is funded by the EPSRC, BBSRC, Wellcome Trust, MRC and a recently awarded ERC Starting Grant. Before starting his group Rui did postdoctoral research in computational neuroscience & machine learning at the University of Oxford and briefly at the University of Bern. Previously, he completed his PhD in 2015 at the University of Edinburgh where he established a collaboration between Mark van Rossum and P. Jesper Sjöström at UCL/McGill.

Event type: Seminar

Date: Tuesday, 23rd January

Time: 11:00

Location: G.03

Speaker(s): Rui Ponte Costa

Chair/Host: David Sterratt