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. We demonstrate that stacking layers and injecting noise during training naturally bias the network towards learning invariant representations. By decomposing the cross-entropy loss, we highlight an inherent overfitting term and propose regularizing it by bounding the information in the weights. This theory enables us to quantify and predict sharp phase transitions between underfitting and overfitting, shedding light on the relation between the geometry of the loss function, invariance, and generalization.
The Paradox of Deep Learning
Deep neural networks present a fundamental theoretical paradox. They are typically "over-parametrized," possessing far more parameters than available training data points. Traditional statistical learning theory suggests such models should suffer from massive overfitting, yet in practice, they generalize remarkably well. They consistently learn "good" representations—features that ignore irrelevant details (nuisances like lighting or background) and focus solely on the task at hand.
This paper addresses two central questions:
- The Invariance Riddle: Why do networks learn to be invariant to transformations like translations or rotations without explicit programming?
- The Generalization Riddle: Why do these massive models generalize to new data instead of simply memorizing the training set?
Information Bottleneck for Weights
The authors bridge the gap between Information Theory and Deep Learning by applying the Information Bottleneck (IB) principle not to the activations, but to the weights themselves. They propose a regularized loss function that penalizes the amount of information the weights $w$ carry about the dataset $\mathcal{D}$.
A key method for limiting weight information is the injection of multiplicative noise during training, a process similar to Variational Dropout. By decomposing the standard cross-entropy loss, the researchers identified four distinct terms: intrinsic error, sufficiency (task information captured), efficiency, and a negative "overfitting" term. By adding a term to bound this overfitting, they derived a theoretically grounded regularization method that relates to a PAC-Bayes perspective.
Minimality and the Invariance Riddle
A central "aha moment" of the research is the proof that Minimality $\equiv$ Invariance. For a representation to be "minimal"—containing the least information necessary to solve a task—it must be invariant to all nuisances. In other words, to forget irrelevant details, the network must achieve a minimal representation.
Furthermore, the paper connects the concept of "flat minima" to Information Theory. A flat minimum in the loss landscape, where weights can be perturbed without significantly changing the output, is simply a weight configuration that requires fewer bits to describe. Stacking noisy layers naturally biases the network toward these low-information, invariant representations, explaining the inherent advantage of "deep" architectures over "wide" ones.
Predicting Overfitting: The $\beta=1$ Phase Transition
The theory predicts a sharp "phase transition" when training on random labels. The regularization strength $\beta$ governs the information complexity. If $\beta > 1$, the network suddenly loses the ability to memorize random noise, and training accuracy drops to chance levels.
As shown in the figure above, the transition between overfitting and underfitting is remarkably sharp. In experiments with CIFAR-10 using random labels, models followed this theoretical prediction regardless of the specific architecture (e.g., AlexNet vs. ResNet). This demonstrates that "Information in Weights" is a far more accurate measure of a model's complexity than a simple parameter count. A model with millions of parameters can have low complexity if its weights reside in a flat minimum or are subject to sufficient noise.
Empirical Validation and Visualizing "Forgetting"
The authors validated their findings across several datasets:
- MNIST: Modified with "clutter" to measure how effectively the network learns to ignore nuisances.
- CIFAR-10: Used for standard classification and the random label experiments mentioned above.
- CelebA: Used to visualize representations through GAN-based reconstructions. As weight information was restricted, the internal representations discarded background details and lighting while strictly preserving the facial features necessary for identity classification.
Conclusion and Future Work
This work establishes a fundamental link between the information stored in a network's weights and the quality of its internal representations. While the theory is elegant, practical challenges remain. Calculating the exact information in weights or the Hessian for flatness is computationally expensive for massive models, requiring proxies like Variational Dropout. Additionally, the current theory focuses on single-task optimization; how these minimal representations perform in transfer learning scenarios remains an open question for future research.
Summary updated