Learning to Compress: Local Rank and Information Compression in Deep Neural Networks

Niket Nikul Patel, Ravid Shwartz-Ziv

TL;DR

This paper introduces "local rank" as a new metric to quantify the dimensionality of learned feature representations in deep neural networks. They show that this local rank decreases during the later stages of training, indicating that networks inherently compress information. This finding connects geometric properties of neural representations to the Information Bottleneck theory, offering a new perspective on how networks learn and generalize.

What Problem Does This Paper Address?

Deep neural networks are known to learn complex representations of data. A key observation is that these representations often lie on low-dimensional "manifolds" within the high-dimensional input space. Understanding how networks learn these low-dimensional representations and why this is beneficial for generalization and robustness is a central question. Existing theories, like the Information Bottleneck (IB) theory, suggest that networks balance compressing irrelevant information while retaining relevant information for the task. However, directly measuring this compression and its geometric underpinnings in learned representations has been challenging. This paper aims to bridge this gap by introducing a concrete measure – local rank – to quantify the dimensionality of these learned feature manifolds and connect it to information compression.

Methods

The core contribution is the introduction of local rank. This metric quantifies the dimensionality of the feature manifold at a specific layer of a neural network. It's defined as the expected rank of the Jacobian of the layer's pre-activation function with respect to the input. To make this computationally tractable, they use a "robust local rank" which considers the rank of the Jacobian matrix after thresholding singular values smaller than a small epsilon ($\epsilon$).

The authors then provide theoretical analysis, showing that under certain optimization conditions (like gradient flow for classification or minimum norm solutions for regression), the local rank is bounded and tends to decrease. This theoretical work is grounded in existing literature on implicit regularization in deep learning, which suggests that standard optimization algorithms (like gradient descent) inherently favor solutions with desirable properties, such as low-rank structures.

Empirically, they train MLPs on synthetic Gaussian data and real-world datasets (MNIST) and measure the local rank across layers during training. They also explore the connection to the Information Bottleneck theory by training Deep Variational Information Bottleneck (VIB) models and observing how local rank changes with the IB trade-off parameter $\beta$.

Datasets

  • Synthetic Gaussian Data: Used to test the theoretical insights in a controlled environment where the underlying data distribution is known.
  • MNIST Dataset: A standard benchmark for image classification, used to demonstrate the empirical findings on a real-world, complex dataset.
  • Fashion-MNIST Dataset: Another image classification dataset, also used to validate the findings.

Key Insights and Ideas

  • Local Rank as a Dimensionality Measure: The introduction of "local rank" provides a concrete, measurable way to quantify the intrinsic dimensionality of the feature manifolds learned by neural networks at each layer. This moves beyond qualitative descriptions of "low-dimensional representations."
  • Rank Reduction During Training: The paper demonstrates that the local rank of representations consistently decreases during the terminal phase of training. This suggests that as networks converge to a solution, they actively compress the dimensionality of their learned features.
  • Connection to Information Bottleneck: The observed rank reduction strongly correlates with information compression, aligning with the Information Bottleneck theory's principle of balancing compression and prediction. This provides a geometric interpretation for the information-theoretic concepts.
  • Implicit Regularization's Role: The theoretical analysis suggests that the observed rank reduction is a consequence of implicit regularization inherent in gradient-based optimization, which favors low-rank solutions.

Key Results

  • Theoretical Bounds: The paper derives theoretical bounds on the local rank, showing it can be related to the norms of weight matrices and the network depth. These bounds suggest that rank reduction is a natural outcome of minimizing network norms.
  • Empirical Rank Drop: Figure 1 clearly shows a significant drop in local rank across all layers during the later stages of training for both synthetic and MNIST datasets. This drop is observed to occur in two phases, consistent with previous findings on information compression in the IB theory.
  • VIB Experiments: Experiments with Deep VIB models (Figure 2 and 3) show that increasing the IB trade-off parameter $\beta$ (which favors compression) leads to an increase in local rank and a decrease in accuracy. Conversely, decreasing $\beta$ (favoring prediction) leads to lower local rank and higher accuracy, reinforcing the link between local rank and information compression.

Limitations and Future Work

  • Non-Gaussian Settings: The theoretical analysis of the Information Bottleneck connection is primarily derived for Gaussian variables. Formalizing this relationship for non-Gaussian settings remains an open area.
  • Network Architectures: The current analysis focuses on Multilayer Perceptrons (MLPs). Extending the concept of local rank and its implications to other architectures like Convolutional Neural Networks (CNNs) or Transformers would be valuable.
  • Practical Applications: While the findings offer insights into model compression and generalization, exploring practical applications of local rank for tasks like model compression or understanding architectural choices is a promising future direction.

Summary updated

Abstract

Deep neural networks tend to exhibit a bias toward low-rank solutions during training, implicitly learning low-dimensional feature representations. This paper investigates how deep multilayer perceptrons (MLPs) encode these feature manifolds and connects this behavior to the Information Bottleneck (IB) theory. We introduce the concept of \emph{local rank} as a measure of feature manifold dimensionality and demonstrate, both theoretically and empirically, that this rank decreases during the final phase of training. We argue that networks that reduce the rank of their learned representations also compress mutual information between inputs and intermediate layers. This work bridges the gap between feature manifold rank and information compression, offering new insights into the interplay between information bottlenecks and representation learning.
BibTeX
@inproceedings{patel2024learning,
  author = {Niket Nikul Patel and Ravid Shwartz-Ziv},
  title = {{Learning to Compress: Local Rank and Information Compression in Deep Neural Networks}},
  booktitle = {Compression @ NeurIPS},
  year = {2024},
  url = {https://openreview.net/forum?id=31LwSPU20u},
}
Figure