Comparative Study between Biological and Artificial Neural Networks
Our research shows that the biological brain organizes visual circuits more efficiently than artificial neural networks. The brain uses fewer, shallower hierarchical networks for visual object recognition, unlike deep neural networks in AI. We aim to explore anatomical structures in the brain that enable this efficiency but are absent in artificial networks.
By comparing brain functions with AI models, can we identify key differences that could improve AI systems and offer new directions for biology-inspired computer vision?
Robust Artificial Visual Intelligence Model of Biological Visual System
Our research examines the limitations of traditional deep neural networks, which struggle with domain-specific biases and visual variations, unlike the adaptable biological brain. Our models maintain robust performance across environmental shifts, aligning with human visual perception. This biologically inspired approach aims to enhance AI models for reliable performance in diverse conditions.
The next question is: How can we further mimic more complex aspects of biological vision? What additional brain mechanisms can be incorporated to enhance AI's robustness across even more diverse and dynamic environments?
Random Pretraining of AI Inspired by Developmental Neuroscience
Unlike AI systems that rely solely on data for training, the biological brain begins learning through random neural activity before sensory input. Inspired by developmental neuroscience, we propose random noise pretraining for AI models. Our research shows that this approach improves learning efficiency and generalization in biologically plausible networks using local learning, and enhances uncertainty calibration in backpropagation-based networks. This highlights how developmental neuroscience insights can address resource efficiency and reliability challenges in AI.
How can these neuroscience-inspired insights contribute further to solving other challenges in AI, such as adaptability, energy efficiency, and robustness?