TL;DR
This paper introduces "local rank" as a new metric to quantify the dimensionality of feature manifolds learned by neural networks. They show, both theoretically and empirically, that this local rank decreases during the final stages of training, indicating that networks compress their internal representations. This finding bridges the gap between geometric properties of learned representations and information-theoretic concepts like the Information Bottleneck, offering a new lens on how deep networks learn.
What Problem Does This Paper Address?
Deep neural networks, especially Multilayer Perceptrons (MLPs), have been observed to implicitly learn low-dimensional representations of data, a phenomenon often referred to as a "bottleneck structure." While this suggests a form of compression, the precise geometric and information-theoretic underpinnings of this process are not fully understood. Specifically, how the dimensionality of the learned feature manifolds evolves during training and how this relates to the network's ability to compress information (as described by the Information Bottleneck theory) remains an open question. This paper aims to bridge this gap by introducing a new metric, "local rank," to quantify this dimensionality and connect it to information compression.
Methods
The core contribution is the introduction and analysis of local rank. This metric quantifies the dimensionality of the feature manifold at a specific layer of a neural network by measuring the expected rank of the Jacobian of the layer's pre-activation function with respect to the input. The authors provide both theoretical analysis and empirical validation:
- Theoretical Analysis: They connect local rank to implicit regularization, showing that under certain optimization objectives (like minimizing weight norms while satisfying data constraints), the local rank is theoretically bounded and expected to decrease. This is demonstrated for both classification and regression tasks.
- Empirical Validation: They train MLPs on synthetic and real-world datasets and measure the local rank across layers throughout training. They also use Deep Variational Information Bottleneck (VIB) models to explore the relationship between local rank and the Information Bottleneck trade-off parameter ($\beta$).
Datasets
- Synthetic Data: A 3-layer MLP was trained on synthetic Gaussian data with a random cross-covariance matrix.
- MNIST Dataset: A 4-layer MLP was trained on the MNIST handwritten digit dataset.
- Fashion-MNIST Dataset: Used in conjunction with MNIST for experiments with Deep VIB models.
Key Insights and Ideas
- Local Rank as a Dimensionality Measure: The paper introduces "local rank" as a principled way to measure the intrinsic dimensionality of the feature manifolds learned by neural networks at each layer. This provides a concrete, quantifiable geometric property.
- Rank Reduction as Compression: The most significant finding is that local rank consistently decreases during the terminal phase of training. This empirically confirms that networks actively compress the dimensionality of their learned representations as they converge, aligning with the idea of an emergent bottleneck.
- Bridging Geometry and Information Theory: By connecting local rank to the Information Bottleneck (IB) framework, the paper offers a new perspective. They show that a reduction in local rank correlates with increased information compression (lower mutual information between input and representation), suggesting that geometric compression (rank reduction) is a mechanism for information compression.
- Phase Transitions in IB: Experiments with Deep VIB models on Gaussian data reveal distinct "phase transitions" in local rank as the IB trade-off parameter ($\beta$) is varied. This suggests that the network's compression behavior is not always smooth and can exhibit critical points.
Key Results
- Theoretical Bounds: The paper provides theoretical bounds on the local rank, showing it is related to the norms of weight matrices and the network depth. These bounds suggest that rank reduction is an inherent consequence of certain optimization strategies.
- Empirical Rank Drop: Figure 1 clearly shows a significant drop in local rank across all layers during the final stages of training for both synthetic and MNIST datasets. This drop is attributed to the second phase of compression identified in IB theory.
- Correlation with IB Parameter: Experiments with Deep VIB models demonstrate that increasing the IB trade-off parameter ($\beta$) leads to an increase in local rank and a decrease in accuracy. This reinforces the idea that higher $\beta$ values encourage less compression and potentially less relevant representations.
Limitations and Future Work
- Non-Gaussian Settings: The theoretical connection between local rank and the Information Bottleneck is explicitly derived for jointly Gaussian variables. The paper acknowledges that formalizing this relationship for non-Gaussian settings is a direction for future research.
- Network Architectures: The current analysis primarily focuses on MLPs. Extending the concept of local rank and its implications to other architectures like Convolutional Neural Networks (CNNs) or Transformers is a natural next step.
- Practical Applications: The authors suggest exploring practical applications of these findings in model compression techniques, leveraging the understanding of how networks naturally compress information through rank reduction.
Summary updated