Neuro-constrained Bio-inspired Recurrent Neural Network

Neural population activity underpins brain function, from sensory processing to decision-making, yet computational models often oversimplify the brain’s complexity. Traditional RNNs, widely used to emulate neural dynamics, assume uniform or random connectivity, neglecting the true synaptic connectivity of neurons (Perich Rajan, 2020). This discrepancy limits the ability of traditional RNNs to accurately replicate behavior, prompting the need for models that integrate anatomical data. Achterberg et al. pioneered the integration of spatial constraints into recurrent neural networks (RNNs) to bridge the gap between artificial models and biological neural systems. Their work demonstrated that spatially embedded RNNs (seRNNs) exhibit enhanced task performance and biologically plausible connectivity patterns. In this work, by integrating anatomical and functional information from biological neurons into an RNN, we aimed to create a model that better reflects the brain’s architecture, offering insights into how structure shapes function. Our approach involved extracting and preprocessing data from the Machine Intelligence from Cortical Networks (MICrONS) dataset, constraining the RNN with anatomical and functional features, and evaluating its performance against baseline models. Below, we detail the datasets, tools, theoretical frameworks, and analytical methods employed.