Reconstruction and causal perturbation of cortical dynamics with recurrent neural networks

This PhD work applies methods from computational and statistical physics to model the dynamics of neural populations underlying action perception. A biologically constrained recurrent neural network was developed to reproduce single-neuron firing activity recorded from macaque cortical areas AIP, F5, and F6 during both action execution and observation in a Go/No-Go reaching task. The model was trained on experimental firing-rate data while enforcing anatomical constraints, including Dale’s principle and restricted local inhibitory projections. The trained networks accurately reconstructed neural activity and revealed an effective connectivity structure characterized by dominant local inhibition, sparse inter-areal coupling, and hierarchical motifs. Crucially, the framework enabled neuron-level perturbations not feasible experimentally, allowing causal exploration of specific cell classes. In silico silencing experiments showed that inhibitory interneurons strongly influenced network stability, particularly in agent discrimination. When examining excitatory populations, non-mirror neurons in F5 and F6 contributed to self-action representation, whereas mirror neurons in AIP preferentially encoded observed actions. Combining experimental data with physics-based modeling of complex dynamical systems, this work establishes a mechanistic framework for investigating the causal architecture of the action observation network and its role in social and motor representations.