Structural Pruning

This study explores the teacher-student paradigm in machine learning, focusing on overparameterized student networks trained by fixed teacher networks. It introduces a new optimization scheme using spectral representation of linear information transfer between layers. This approach allows identifying a stable student substructure that mirrors the teacher’s complexity. The method shows that pruning unimportant nodes, based on optimized eigenvalues, does not degrade performance, indicating a second-order phase transition with universality traits in neural network training.

Training neural networks in spectral space focuses on optimizing eigenvalues and eigenvectors instead of individual weights, allowing effective implicit bias that node enables pruning without sacrificing performance.

This study presents a variant of spectral learning for deep neural networks, where adjusting two sets of eigenvalues for each layer mapping significantly enhances network performance with fewer trainable parameters. This method, inspired by homeostatic plasticity, offers a computationally efficient alternative to conventional training, achieving comparable results with a simpler parameter setup. It also enables the creation of sparser networks with impressive classification abilities.