Fig. 5: MD context inference with more complex tasks and PFC learning.
From: Rapid context inference in a thalamocortical model using recurrent neural networks
A The unique challenge of rapid context inference with complex task dynamics across trials. Different sets of neurons are activated within and across trials. The trace computation can not guarantee to cover all target task neurons that are required for all possible inputs in specific tasks. B The designed network architecture with two pathways: the task learning pathway optimized with supervised learning through the PFC and the context inference pathway through the PFC-ctx to regulate the PFC neurons. C Comparison of rules decoding and cueing context decoding accuracy in the mice experiment data, measured as mutual information, in the PFC transient neurons (cue-selective and cue invariant), the PFC fast spiking neurons and the MD neurons. ***P<0.001, Bonferroni-corrected Kruskal–Wallis ANOVA. Figure 5C reproduced from Fig. 2F of Rikhye, R.V., Gilra, A. & Halassa, M.M. Thalamic regulation of switching between cortical representations enables cognitive flexibility. (https://doi.org/10.1038/s41593-018-0269-z). D The context decoding performance of different modules: the MD, the PFC-ctx, and the PFC. ***P < 0.001, **P < 0.005, Bonferroni-Corrected Rank-Sum test. E The context decoding performance of different modules: the MD, the PFC-ctx, and the PFC over time within each trial. F The neural trajectories of different tasks in the PFC only model and the PFC-MD model.
