Abstract:An antithetical concept, adaptive symmetry, to conservative symmetry in physics is proposed to understand the deep neural networks (DNNs). It characterizes the invariance of variance, where a biotic system explores different pathways of evolution with equal probability in absence of feedback signals, and complex functional structure emerges from quantitative accumulation of adaptive-symmetries breaking in response to feedback signals. Theoretically and experimentally, we characterize the optimization process of a DNN system as an extended adaptive-symmetry-breaking process. One particular finding is that a hierarchically large DNN would have a large reservoir of adaptive symmetries, and when the information capacity of the reservoir exceeds the complexity of the dataset, the system could absorb all perturbations of the examples and self-organize into a functional structure of zero training errors measured by a certain surrogate risk. More specifically, this process is characterized by a statistical-mechanical model that could be appreciated as a generalization of statistics physics to the DNN organized complex system, and characterizes regularities in higher dimensionality. The model consists of three constitutes that could be appreciated as the counterparts of Boltzmann distribution, Ising model, and conservative symmetry, respectively: (1) a stochastic definition/interpretation of DNNs that is a multilayer probabilistic graphical model, (2) a formalism of circuits that perform biological computation, (3) a circuit symmetry from which self-similarity between the microscopic and the macroscopic adaptability manifests. The model is analyzed with a method referred as the statistical assembly method that analyzes the coarse-grained behaviors (over a symmetry group) of the heterogeneous hierarchical many-body interaction in DNNs.

Do deep neural networks behave like structural glasses?

Comparing Dynamics: Deep Neural Networks versus Glassy Systems

What Do Deep Neural Networks Find in Disordered Structures of Glasses?

Neural Networks as Spin Models: From Glass to Hidden Order Through Training

Towards interpreting the thermally activated $β$ dynamics in metallic glass with the structural constraint neural network

Glassy dynamics near the learnability transition in deep recurrent networks

Toward Interpreting the Thermally Activated β Dynamics in Metallic Glass Using the Structural Constraint Neural Network

Microscopic theory of the glassy dynamics of passive and active network materials

Exploring Loss Landscapes through the Lens of Spin Glass Theory

A structural approach to relaxation in glassy liquids

A geometry-enhanced graph neural network for learning the smoothness of glassy dynamics from static structure

Learning molecular dynamics: predicting the dynamics of glasses by a machine learning simulator

A deep learning approach to the measurement of long-lived memory kernels from Generalised Langevin Dynamics

Rotation-equivariant Graph Neural Networks for Learning Glassy Liquids Representations

Neural annealing and visualization of autoregressive neural networks in the Newman-Moore model

Dynamic heterogeneity at the experimental glass transition predicted by transferable machine learning

Charting the Topography of the Neural Network Landscape with Thermal-Like Noise

Strength of 2D Glasses Explored by Machine-Learning Force Fields

Topological Constraints and Rigidity of Network Glasses from Molecular Dynamics Simulations

Complexity from Adaptive-Symmetries Breaking: Global Minima in the Statistical Mechanics of Deep Neural Networks

Order and Chaos: NTK views on DNN Normalization, Checkerboard and Boundary Artifacts