Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session C: Friday, August 15, 2:00 – 5:00 pm, de Brug & E‑Hall

Controlling PFC dynamics for slow and fast learning

Michał J Wójcik¹, Jascha Achterberg¹, Joseph Oliver Pemberton², Rui Ponte Costa¹; ¹University of Oxford, ²University of Washington

Presenter: Michał J Wójcik

The prefrontal cortex (PFC) exhibits a remarkable capacity to employ two distinct strategies when engaging in cognitive tasks. Upon encountering a novel task, it leverages high-dimensional representations, well positioned for rapid linear decoding. However, with growing task familiarity, the PFC transitions to employing generalisable low-dimensional neural codes. Through a system-level modelling approach, we propose that these properties emerge naturally in recurrent neural networks (RNNs) that learn on two distinct timescales: (i) on a faster timescale an external controller drives RNN dynamics to generate task-encoding but relatively unstructured, high-dimensional representations, which is then followed by (ii) a slower optimisation of recurrent connections and consequently more structured, low-dimensional representations. We validated these predictions by comparing model representations to neural recordings from the prefrontal cortex of non-human primates that were trained to learn a complex cognitive task from scratch. In summary, our results suggest a learning-dependent control of prefrontal dynamics via a separate brain-region for high-to-low representational switching.

Topic Area: Predictive Processing & Cognitive Control

Extended Abstract: Full Text PDF