Neural Network Architecture Generation Driven by Neuron Circuitry

Abstract

This paper proposes a novel approach to neural network architecture generation, termed "Neuron Circuitry Driven Dynamic Program Generation." The core claim is that comprehensive analysis of large-scale neuron circuits can automatically generate highly optimized neural network architectures and connection weights tailored to specific tasks, surpassing the capabilities of traditional manual design or search-based methods. The proposed method leverages deep learning to train multiple large networks, extracting detailed neuron circuit maps (s1, s2, ..., sK) representing neuron connections – including connection strength, direction, and type. Graph Neural Networks (GNNs) are then employed to learn and reason from these circuit maps, identifying optimal connection patterns and architectural features for a given task. The resulting neural network architectures are dynamic, adapting in real-time to input data and task requirements for continuous self-optimization. This approach represents a significant departure from existing methods, which typically rely on static architectures or static search, by directly learning and utilizing biological neuron circuit maps through GNN modeling and inference, offering the potential for substantially improved neural network performance and efficiency.