Function space analysis of deep learning representation layers

Oren Elisha, Shai DekelarXiv 2017 · arXiv

Introduction

Deep Learning (DL) models are often characterized as "black boxes" due to the difficulty in quantifying how hidden layers transform raw data into abstract, separable features. A common intuition is that these layers "uncrumple" a high-dimensional data manifold, sequentially simplifying its geometric structure. This paper formalizes this intuition by introducing a novel metric based on Besov smoothness indices to measure the quality and mathematical smoothness of deep learning representations layer-by-layer.

Methodology: Measuring Smoothness in Function Space

The authors bridge Approximation Theory and Representation Learning by treating each network layer as a functional mapping $f_k$ from input features to a vector-valued representation of class labels. To quantify the smoothness of these mappings without falling victim to the curse of dimensionality, they employ a Wavelet Decomposition of Random Forests (RF).

Key components of their approach include:

  1. Functional Mapping: Representations are mapped to vertices of a simplex representing class labels.
  2. Random Forest Wavelets: A Random Forest is fitted to the data at each layer, followed by a wavelet decomposition to identify significant components of the representation.
  3. Dimension-Free Besov Index ($\alpha$): The authors derive a "weak-type" Besov smoothness index that is dimension-invariant. This allows for direct comparison between layers of vastly different widths (e.g., comparing a 10,000-dimensional input to a 128-dimensional hidden layer).
  4. Numerical Estimation: The index $\alpha$ is estimated by observing the decay rate of approximation error as more wavelet terms are included. A faster decay indicates a smoother, more compressible representation.

Experimental Results

Layer-by-Layer "Uncrumpling"

The primary finding is that in successful networks, the Besov smoothness index monotonically increases from the input layer to the final output. This provides empirical evidence for the "uncrumpling" hypothesis. For instance, in experiments using the Urban8K audio dataset, the smoothness index rose significantly across successive layers.

Layer-by-layer increase in Besov smoothness index for the Urban8K dataset.

As shown in Figure 3, the smoothness index starts at approximately 0.13 at the input and reaches over 0.35 at the final layer, indicating a progressively more structured and "smooth" feature space. Similar trends were observed in MNIST and CIFAR-10 experiments.

Smoothness and Training Dynamics

Smoothness is not just a static property of the architecture but an emergent property of the learning process. The authors observed that the Besov index increases not only across layers but also across training epochs.

Comparison of smoothness indices across layers at 20 and 50 training epochs.

Figure 5 illustrates this progression: as training continues from 20 to 50 epochs, the representations at every layer become mathematically smoother. This suggests that the network is actively refining the geometric structure of the data manifold to improve class separability.

Generalization and Sensitivity to Noise

The study also explored the relationship between smoothness and generalization. By intentionally mis-labeling training data, the authors found that the Besov smoothness plummeted. In MNIST tests, the index dropped from 0.28 (clean data) to 0.03 (40% mis-labeling). This indicates that overfitting to noise results in "jagged," non-smooth feature spaces, making the Besov index a potential proxy for detecting poor generalization.

Conclusion

By treating deep learning layers as transformations in function space, this work provides a rigorous mathematical framework for understanding representation learning. The dimension-free Besov smoothness index offers a powerful tool for quantifying the "uncrumpling" of data manifolds, revealing how both network depth and training duration contribute to the creation of structured, smooth feature representations. Future work may extend this analysis to more complex architectures like Transformers or explore the specific roles of pooling and non-linearities in this smoothing process.

Summary updated

Abstract

In this paper we propose a function space approach to Representation Learning and the analysis of the representation layers in deep learning architectures. We show how to compute a weak-type Besov smoothness index that quantifies the geometry of the clustering in the feature space. This approach was already applied successfully to improve the performance of machine learning algorithms such as the Random Forest and tree-based Gradient Boosting. Our experiments demonstrate that in well-known and well-performing trained networks, the Besov smoothness of the training set, measured in the corresponding hidden layer feature map representation, increases from layer to layer. We also contribute to the understanding of generalization by showing how the Besov smoothness of the representations, decreases as we add more mis-labeling to the training data. We hope this approach will contribute to the de-mystification of some aspects of deep learning.
BibTeX
@misc{elisha2017function,
  author = {Oren Elisha and Shai Dekel},
  title = {{Function space analysis of deep learning representation layers}},
  year = {2017},
  eprint = {1710.03263},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1710.03263},
}
Figure