CWI Scientific Meeting 6 October 2023

Alexandra van den Berg (Machine Learning), Sami Mollard (Machine Learning)

When
6 Oct 2023 from 1 p.m. to 6 Oct 2023 2 p.m. CEST (GMT+0200)
Where
L016
Web
Add

13:00 - 13:30 Alexandra van den Berg (Machine Learning), How the brain can learn to learn

Abstract: Learning does not occur in isolation. Instead, humans accumulate knowledge over time and use their previous experiences to accelerate learning on new problems. This process is called meta-learning or learning-to-learn. This is in contrast to artificial neural networks, which are prone to catastrophic forgetting when task requirements change. While models have been developed that overcome this problem, they typically differ from the brain in important ways. In particular, they learn with backpropagation-through-time (BPTT), which requires information that is not locally available at the synapses undergoing plasticity. In our work, we developed a novel gated recurrent network named RECOLLECT that can flexibly store task-related information in its memory and learn-to-learn in a way that resembles animal learning using only local information for training. As such, we can use RECOLLECT to further our understanding of how learning-to-learn is performed in the brain.

13:30 - 14:00 Sami Mollard (Machine Learning), How the brain learns to solve multistep, multiscale visual tasks

Abstract: Visual information is processed, in primates' brain, in the visual cortex. More than just integrating bottom-up sensory signals coming from the retina, the early visual cortex also receives information from higher cortical area through feedback projections. It has been proposed that those feedback connections enable the visual cortex to act as a cognitive blackboard where task-relevant features are highlighted as a focus of enhanced activity, which is particularly useful to solve multistep, multiscale visual tasks. However, it remains unclear how such strategies are learnt and implemented. To answer this question, we developed recurrent neural networks trained with reinforcement learning to solve such tasks, on which monkeys were also trained. Interestingly, after training, the patterns of activations of artificial and biological neurons were similar. Our results shed light on how multiscale, multistep visual tasks are learnt in the visual cortex.