Jeffrey's prior sampling of deep sigmoidal networks

Lorien X. Hayden, Alexander A. Alemi, Paul H. Ginsparg, James P. SethnaarXiv 2017 · OpenReview

TL;DR

This paper investigates the geometric structure of the "model manifold" learned by deep neural networks (specifically DBNs and SdAs) trained for image reconstruction. Instead of the expected "hyper-ribbon" structure seen in other complex models, they find these neural networks learn a more compact, "slightly elongated hyperball." This suggests that the data learned by these networks primarily resides on the boundaries of this manifold, raising questions about the efficiency of their learned representations.

What Problem Does This Paper Address?

Deep neural networks are known for their ability to learn complex, low-dimensional representations of high-dimensional data. When trained for tasks like image reconstruction, the set of all possible outputs forms a "model manifold." Understanding the geometry of this manifold is crucial for interpreting what the network has learned and how it represents data. Previous work on other complex models (like those in systems biology or statistical physics) has often shown these manifolds to have a "sloppy" or "hyper-ribbon" structure, where certain parameter combinations have a disproportionately large effect on the output. This paper asks: does this hyper-ribbon structure also apply to deep neural networks, and what does the geometry of their learned manifolds tell us about their internal workings?

Methods

The authors employ Monte Carlo sampling, specifically Metropolis-Hastings, to explore the model manifold. A key methodological contribution is the use of Jeffrey's prior, which is derived from the Fisher Information Matrix. This prior is chosen because it weights parameter space by data space volume, aiming to avoid getting stuck in regions of low data probability. They analyze the geometry of the learned manifold by examining its shape and the distribution of actual data points within it, often using Principal Component Analysis (PCA) to visualize projections of the high-dimensional manifold.

Datasets

The primary dataset used is the MNIST dataset of handwritten digits. They train three types of networks on this dataset:

  • A Stacked Denoising Autoencoder (SdA) trained on a single digit ('1').
  • A Stacked Denoising Autoencoder (SdA) trained on the full MNIST dataset.
  • A Deep Belief Network (DBN) trained on the full MNIST dataset.

Key Insights and Ideas

  • Unexpected Geometry: Deep neural networks trained for reconstruction do not form the expected "hyper-ribbon" manifold. Instead, they learn a more compact, "slightly elongated hyperball." This suggests that the parameters in these networks might be weighted more equally in their influence on the output, unlike the "sloppy" models seen elsewhere.
  • Data on the Boundaries: A significant finding is that the actual MNIST data points (and their reconstructions) tend to lie on the boundaries of the learned model manifold, rather than in its interior. This is particularly evident in the PCA projections.
  • Jeffrey's Prior and Exploration: Using Jeffrey's prior for sampling helps explore the manifold, but the authors note that it might bias exploration towards regions that are not necessarily the most "interesting" or informative, especially given the data's tendency to reside on the boundaries.
  • Sigmoidal vs. Linear Layers: The behavior of the manifold's structure across different layers seems to be influenced by whether the layers are sigmoidal or linear. Sigmoidal layers appear to contribute to the boundary-hugging behavior.

Key Results

  • PCA projections of the sampled model manifolds reveal an "only slightly elongated hyperball" for all networks studied (DBN, SdA, and single-digit '1' SdA).
  • The width of the manifold shrinks by a factor of 1.17 along PCA directions for the 10-dimensional top layer of the single-digit network, and a much smaller factor of 1.03 for the 30-dimensional top layers of the DBN and SdA. This contrasts sharply with the "hyper-ribbon" structures observed in other models.
  • Actual MNIST digit data and their reconstructions consistently appear on the edges or boundaries of the sampled model manifolds, even when visualized in lower-dimensional PCA projections.
  • The paper presents visualizations (Figures 3, 7, 8, 10, 12, 13) showing these geometric properties, including PCA projections of the data and sampled points, and examples of reconstructed and sampled images.

Limitations and Future Work

  • Interpretation of Boundary Data: The authors acknowledge that the saturated pixels in the MNIST dataset (due to sigmoidal activation functions) might contribute to the data appearing on the manifold's boundaries, making it difficult to definitively interpret this as a general property of all deep networks or all datasets.
  • Jeffrey's Prior Bias: They suggest that while Jeffrey's prior is useful, it might be biased towards exploring uninteresting regions of the manifold, especially when the data itself is concentrated on the boundaries. This raises the question of whether the network is "wasteful" in its representation.
  • Future Work: The paper implicitly suggests future work could involve exploring different datasets, network architectures (e.g., different activation functions), and sampling priors to better understand the relationship between data, network structure, and manifold geometry. The observation that lower-dimensional representations also hug the outside of the manifold suggests further investigation into hierarchical representations.

Summary updated

Abstract

Neural networks have been shown to have a remarkable ability to uncover low dimensional structure in data: the space of possible reconstructed images form a reduced model manifold in image space. We explore this idea directly by analyzing the manifold learned by Deep Belief Networks and Stacked Denoising Autoencoders using Monte Carlo sampling. The model manifold forms an only slightly elongated hyperball with actual reconstructed data appearing predominantly on the boundaries of the manifold. In connection with the results we present, we discuss problems of sampling high-dimensional manifolds as well as recent work [M. Transtrum, G. Hart, and P. Qiu, Submitted (2014)] discussing the relation between high dimensional geometry and model reduction.
BibTeX
@misc{hayden2017jeffrey,
  author = {Lorien X. Hayden and Alexander A. Alemi and Paul H. Ginsparg and James P. Sethna},
  title = {{Jeffrey's prior sampling of deep sigmoidal networks}},
  year = {2017},
  eprint = {1705.10589},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1705.10589},
}
Figure