A new way to build neural networks could make AI more understandable.

Ananthaswamy, A. (2024, August 29). A new way to build neural networks could make AI more understandable. MIT Technology Review. https://physics.mit.edu/news/a-new-way-to-build-neural-networks-could-make-ai-more-understandable/

This report from MIT Physics details the emergence of Kolmogorov-Arnold Networks (KANs), a reimagined neural network architecture designed to solve the "black box" problem of modern artificial intelligence. Unlike traditional models that use fixed internal activation functions, KANs employ learnable functions moved outside the neuron, allowing the system to be much more transparent and interpretable to human researchers. By simplifying the mathematical structure of the individual building blocks, these networks enable scientists to reverse-engineer decisions and identify the specific formulas the AI uses to reach a conclusion. While currently more computationally expensive than standard methods, KANs demonstrate superior accuracy and scalability in scientific tasks, potentially transforming AI from an inscrutable tool into a verifiable and understandable partner in research.