Where is the Information in a Deep Neural Network?

Abstract

Whatever information a deep neural network has gleaned from training data is encoded in its weights. How this information affects the response of the network to future data remains largely an open question. Indeed, even defining and measuring information entails some subtleties, since a trained network is a deterministic map, so standard information measures can be degenerate. We measure information in a neural network via the optimal trade-off between accuracy of the response and complexity of the weights, measured by their coding length. Depending on the choice of code, the definition can reduce to standard measures such as Shannon Mutual Information and Fisher Information. However, the more general definition allows us to relate information to generalization and invariance, through a novel notion of effective information in the activations of a deep network. We establish a novel relation between the information in the weights and the effective information in the activations, and use this result to show that models with low (information) complexity not only generalize better, but are bound to learn invariant representations of future inputs. These relations hinge not only on the architecture of the model, but also on how it is trained, highlighting the complex inter-dependency between the class of functions implemented by deep neural networks, the loss function used for training them from finite data, and the inductive bias implicit in the optimization.

References in corpus (17)

• Striving for Simplicity: The All Convolutional Net
• The information bottleneck method
• Opening the Black Box of Deep Neural Networks via Information
• On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima
• Variational Dropout and the Local Reparameterization Trick
• Deep Variational Information Bottleneck
• Train faster, generalize better: Stability of stochastic gradient descent
• New insights and perspectives on the natural gradient method
• Understanding disentangling in $\beta$-VAE
• Information-theoretic analysis of generalization capability of learning algorithms
• Sharp Minima Can Generalize For Deep Nets
• A PAC-Bayesian Tutorial with A Dropout Bound
• Information-Theoretic Generalization Bounds for SGLD via Data-Dependent Estimates
• Chaining Mutual Information and Tightening Generalization Bounds
• Dynamics and Reachability of Learning Tasks
• Quantitative Weak Convergence for Discrete Stochastic Processes
• On the Convex Behavior of Deep Neural Networks in Relation to the Layers' Width

Cited by in corpus (17)

• Deep Neural Networks and Tabular Data: A Survey
• Social physics
• Fast Yet Effective Machine Unlearning
• Deep Learning Theory Review: An Optimal Control and Dynamical Systems Perspective
• A Diffusion Theory For Deep Learning Dynamics: Stochastic Gradient Descent Exponentially Favors Flat Minima
• Generalized Variational Continual Learning
• Layer-wise Characterization of Latent Information Leakage in Federated Learning
• Information in Infinite Ensembles of Infinitely-Wide Neural Networks
• Artificial Neural Variability for Deep Learning: On Overfitting, Noise Memorization, and Catastrophic Forgetting
• Improving Generalization by Controlling Label-Noise Information in Neural Network Weights
• Inversion dynamics of class manifolds in deep learning reveals tradeoffs underlying generalisation
• Training Neural Networks with Fixed Sparse Masks
• Estimating informativeness of samples with Smooth Unique Information
• The distance between the weights of the neural network is meaningful
• An Extension of Fano's Inequality for Characterizing Model Susceptibility to Membership Inference Attacks
• Whitening and second order optimization both make information in the dataset unusable during training, and can reduce or prevent generalization
• Quantifying the effect of representations on task complexity