Institute for Basic Science Develops Lp-Convolution: Brain-Inspired AI Vision Breakthrough

Institute for Basic Science (IBS), Yonsei University, and Max Planck Institute researchers have developed Lp-Convolution, a revolutionary AI technique that brings machine vision closer to how the human brain processes images, marking a significant advancement in both AI and neuroscience.

Published on April 22, 2025, this breakthrough addresses a fundamental limitation in current AI systems. While the human brain efficiently identifies key details in complex scenes, traditional Convolutional Neural Networks (CNNs) struggle with their rigid, square-filter approach. Vision Transformers (ViTs) offer superior performance but require massive computational power.

Lp-Convolution uses a multivariate p-generalized normal distribution (MPND) to reshape CNN filters dynamically. Unlike traditional CNNs with fixed square filters, this method allows AI models to adapt filter shapes—stretching horizontally or vertically based on the task, mimicking how the brain selectively focuses on relevant details.

This breakthrough solves the "large kernel problem" where simply increasing filter sizes in CNNs typically doesn't improve performance despite adding parameters. Lp-Convolution overcomes this through flexible, biologically inspired connectivity patterns.

Tests on standard image classification datasets (CIFAR-100, TinyImageNet) showed significant accuracy improvements on both classic models like AlexNet and modern architectures like RepLKNet. The method proved highly robust against corrupted data, crucial for real-world applications.

When Lp-masks resembled Gaussian distributions, AI processing patterns closely matched biological neural activity, confirmed through mouse brain data comparisons.

Dr. C. Justin LEE, Director of the Center for Cognition and Sociality, emphasized: "Our Lp-Convolution mimics the brain's ability to flexibly focus on relevant image parts, making AI smarter and more adaptable."

This innovation could revolutionize autonomous driving (real-time obstacle detection), medical imaging (enhanced AI-based diagnoses), and robotics (adaptable machine vision under changing conditions).

The research will be presented at ICLR 2025, with code and models publicly available, representing a powerful contribution to both AI and neuroscience by aligning artificial systems more closely with biological intelligence.