Abstract
Using established principles from Statistics and Information Theory, we show that invariance to nuisance factors in a deep neural network is equivalent to information minimality of the learned representation, and that stacking layers and injecting noise during training naturally bias the network towards learning invariant representations. We then decompose the cross-entropy loss used during training and highlight the presence of an inherent overfitting term. We propose regularizing the loss by bounding such a term in two equivalent ways: One with a Kullbach-Leibler term, which relates to a PAC-Bayes perspective; the other using the information in the weights as a measure of complexity of a learned model, yielding a novel Information Bottleneck for the weights. Finally, we show that invariance and independence of the components of the representation learned by the network are bounded above and below by the information in the weights, and therefore are implicitly optimized during training. The theory enables us to quantify and predict sharp phase transitions between underfitting and overfitting of random labels when using our regularized loss, which we verify in experiments, and sheds light on the relation between the geometry of the loss function, invariance properties of the learned representation, and generalization error.
Introduction
Deep neural networks are notoriously over-parameterized, often possessing far more weights than training samples. According to classical statistical theory, such models should suffer from extreme overfitting; however, in practice, they generalize remarkably well. This paper bridges two parallel explanations for this phenomenon: the optimization perspective, which focuses on the "flatness" of the loss landscape, and the representation learning perspective, which emphasizes the network's ability to learn features invariant to "nuisances" (irrelevant data variations).
By establishing a rigorous information-theoretic link between these views, the authors demonstrate how the process of optimizing weights (the network's "past") leads to desirable properties in the activations (the network's "future" inference on new data).
The Information Bottleneck for Weights
The core methodology involves a novel application of the Information Bottleneck (IB) framework. While traditional IB focuses on the bottleneck between input and activations, this work introduces an Information Bottleneck for the Weights.
Key components of this approach include:
- Loss Decomposition: The standard cross-entropy loss is decomposed into four terms: intrinsic error, sufficiency (task-relevant information), efficiency, and a negative "overfitting" term representing memorization.
- Weight Information Regularization: The authors propose regularizing the information the weights $w$ contain about the dataset $\mathcal{D}$, denoted as $I(w; \mathcal{D})$.
- Stochastic Weights: To measure and limit information, weights are treated as random variables by injecting noise during training, a technique similar to Variational Dropout.
Theoretical Insights: Invariance and Flat Minima
The paper's primary theoretical contribution is the proof that Minimality = Invariance. If a representation is sufficient to solve a task but contains the absolute minimum information about the input, it is mathematically forced to be invariant to nuisance factors.
Furthermore, the authors link this to the geometry of the loss landscape. "Flat minima"—regions where weights can be perturbed without significantly changing the output—are shown to be a proxy for low weight information. Because flat minima cannot store high-precision information, they naturally prevent the network from memorizing the specificities of the training set, thereby promoting generalization.
Experimental Results and Phase Transitions
The theory was validated using benchmarks including MNIST (with random clutter), CelebA (for feature visualization), and CIFAR-10. A key prediction of the theory is the existence of sharp phase transitions in model behavior based on the regularization strength $\beta$.
As shown in the figure above, experiments with random labels reveal a sharp transition at the critical value $\beta = 1$ (indicated by the dashed green line). Below this value, the network has enough capacity to memorize random labels (overfitting), while above it, the network suddenly stops memorizing and begins to underfit. In the case of real labels, a "fitting" region emerges where the network successfully generalizes, achieving high testing accuracy by balancing task sufficiency with weight minimality.
Conclusion
This work provides a unified framework for understanding generalization in deep learning. By showing that invariance and disentanglement are bounded by the information in the weights, it explains how standard training practices—like stacking layers and using SGD to find flat minima—implicitly optimize for robust representations. While the current theory relies on a stochastic view of networks, it offers a powerful tool for quantifying model complexity and predicting the limits of learning.
Summary updated