A Novel Quantum-based Model of the Cortical Canonical Microcircuit

Overview

Most complex systems are made up of simple, repeating functional units. There has been great controversy over whether such a unit, a ‘canonical microcircuit’, makes up the cortex of the human brain. Although abundant evidence demonstrates uniformity in str ucture and function across the cortex, determining the exact computation performed by an individual microcircuit has proved elusive. The leading proposal for this ‘universal computation’ is Divisive Normalization, which only achieves a FEV (fraction of explainable variance explained) score of 52% in predicting the behavior of neuron populations in the brain. In this project, a Quantum Convolutional Neural Network (QCNN) model is proposed as the canonical computation for a variety of reasons: a) parallel computation is performed by design, b) energetically expensive and time-consuming physical information exchange across neurons is obviated by the entangling nature intrinsic to qubits, and c) higher explainability, parsimony, and applicability to biological systems is achieved via quantum circuit modeling rather than a series of matrix multiplications. The model was trained and tested using Qiskit on 40x40 images of synthetic and natural stimuli, and the predicted vs. actual firing rates were measured. Further datasets (MNIST, Fashion -MNIST, and Superimposed Noise Datasets) were used to test feature extraction properties and robustness of the model, as well as validate the accuracy. QCNN consistently outperformed classical models in FEV, making it an attractive candidate for canonical computation.