Deep neural networks, as a core technology in the field of artificial intelligence, have always been a key research focus in the scientific community due to their complex operating mechanisms and performance optimization. The team led by Professor Ivan Dokmanić from the Department of Mathematics and Computer Science at the University of Basel has recently achieved a breakthrough by constructing an intuitively understandable mechanical model that successfully reproduces key features of deep neural networks, providing new ideas for optimizing network parameters. The related research results have been published in Physical Review Letters.

The model developed by the team is based on a folding ruler as a prototype, with its various parts corresponding to different layers of the neural network. In the experiments, the researchers simulated nonlinear computations and random noise in neural networks by controlling the pulling speed and shaking amplitude. Dr. Cheng Shi explained: “When pulled slowly, the folding ruler only unfolds at the front end, corresponding to data separation concentrated in shallow layers; when pulled quickly and shaken, all parts unfold evenly, simulating the balanced distribution of data across layers.” This discovery verifies the importance of balanced data separation for improving network performance and provides an intuitive framework for understanding the internal operating mechanisms of neural networks.

In traditional neural network training, the interaction between nonlinear computations and noise often leads to complex behaviors that are difficult to describe mathematically with precision. Drawing inspiration from physical theories, Professor Dokmanić’s team materialized abstract concepts through a mechanical model. Mr. Shi pointed out: “The simulation results show that the behavior of the mechanical model is highly consistent with that of real networks, providing a theoretical basis for optimizing network structures and reducing training resource consumption.” Currently, the team is extending this approach to large language models, and in the future, it may replace traditional trial-and-error methods to achieve precise training of high-performance neural networks.