How a student becomes a teacher: learning and forgetting through Spectral methods
We test the performance of a novel pruning strategy grounded on the inspection of the eigenvalues magnitude at the end of the training. The method is applied to a fully connected network trained on tasks of increasing complexity. When the complexity of the function is known we are capable of pruning the network to a stable minimal substructure, which is able to perform the task with the same accuracy of the full network.Abstract
In theoretical Machine Learning, the teacher-student paradigm is often employed as an effective metaphor for real-life tuition. A student network is trained on data generated by a fixed teacher network until it matches the instructor’s ability to cope with the assigned task. The study proposes a novel optimization scheme building on a spectral representation of linear transfer of information between layers. This method enables the isolation of a stable student substructure, mirroring the teacher’s complexity in terms of computing neurons, path distribution, and topological attributes. Pruning unimportant nodes based on optimized eigenvalues retains full network performance, illustrating a second-order phase transition in neural network training.
Lorenzo Giambagli
PostDoc Department of Physics, Freie Universität Berlin
My research interests include Spectral analysis of Deep Neural Network (DNN), Structura Pruning, Bayesian Inference in DNN, Simplicial Complexes Dynamics, Theoretical Neuroscience