Synergetic Learning Systems: Interpretation of Deep Learning Using Physics
Ping Guo, Yuping Wang, Hai‐Lin Liu · 2025
Deep neural networks (DNN) are widely used in various fields and get great success, but one of the major challenges is interpretable mechanism of DNN model. In order to break through the limitations of current deep learning, especially for the interpretability problem of DNN models, we will give a physics-based interpretation of the DNN model in this work. Due to the relationship among autoencoder, minimum description length, mutual information, Ising model, renormalization group etc and free energy principle in deep learning, the energy-based learning framework provides some solid foundations for interpretable DNN model. Therefore, based on the least action principle, the interpretation for generative models using physics principle is presented, and the interpretability problem of DNN model is investigated. Our proposed Synergetic Learning Systems (SLS) is an artificial intelligence (AI) system with multiple subsystems. Here we take two-model SLS, one is the reductive model and another is evolutionary model, as an example to illuminate the interpretability of the DNN models. This work is significant as it offers a clear physical perspective on interpretable DNNs, paves the road for the development of physical artificial intelligence, and prompts development of Science for AI greatly.