How do we learn?

We are at an exciting turning point in neuroscience. New technologies now allow us to measure and control neural activity and behaviour with unprecedented detail (Landhuis et al. Nature 2017, Lauer et al. Nature Methods 2022). At the same time, new theoretical frameworks are starting to reveal how rich behaviours arise from synaptic, circuit and systems computations (Richards et al. Nature Neuroscience 2019). In our group, we are contributing directly to the latter by aiming to understand how we learn. To this end, we are developing a new generation of computational models of brain function driven by recent machine learning developments.

We focus on understanding how a given behavioural outcome ultimately leads to credit being assigned to trillions of synapses across multiple brain areas – credit assignment problem (Fig. 1a). We believe that in order to have a unified understanding of how we learn to produce adaptable behaviours it is important to jointly study the contribution of three different systems (Fig. 1b): (i) cortical circuits, (ii) neuromodulation and (iii) subcortical regions.

Figure 1: Three key dimensions for credit assignment in the cortex. (a) Learning requires the sensory stimuli to be associated with specific behaviours. For successful learning the brain should assign credit to trillions of synapses given behavioural outcomes – synaptic credit assignment problem. (b) We propose that it is critical to jointly study how (i) neocortical circuits (gray), (ii) subcortical systems (green) and (iii) neuromodulation (orange) shape credit assignment to make substantial progress in our understanding of how animals learn naturalistic and complex tasks. This research program is closely guided by both machine learning and experimental findings.

Cortical circuits for efficient credit assignment

We are interested in understanding what are the principles that enable cortical circuits to learn non-trivial problems efficiently. We establish close links between state-of-the-art deep learning algorithms and circuits across the cortex, thereby providing new prespectives on how these circuits ultimately lead to adaptive behaviour.

Neuromodulation of cortical credit assignment

Neuromodulation is critical to establish bridges between internal representations of the world and feedback from the enviroment. We are particularly interested in understanding how the classical source of reward prediction errors (dopaminergic system) and the cholinergic system jointly control cortex-wide learning processes. This work is already providing novel theories about the underlying causes of cognitive decline in dementia, ageing and injury.

Subcortical regions as facilitators of cortical credit assignment

Cortical and subcortical structures have evolved jointly. This strongly suggests that these two must cooperate to enable adaptive behaviour. Recently, we have introduced novel theories of how the mini-brain (cerebellum) may unlock learning in the cortex by predicting future feedback. We are also modelling the hippocampus and are interested in understanding how it guides the cortex during learning and planning.

Neural & ML       Neural & ML       first.verylast (at)

Made with Jekyll based on Kording Lab template.