The 'tree-dependent components' of natural scenes are edge filters

Daniel Zoran, Yair WeissNeurIPS 2009 · OpenReview

Most traditional models of natural image statistics attempt to force image components to be entirely independent. However, natural images are stubbornly full of dependencies. In their paper, researchers Daniel Zoran and Yair Weiss flip this script: instead of minimizing dependency, they explicitly maximize a tree-structured dependency. By doing so, they discover that this approach naturally produces edge filters and mimics the "complex cell" behavior found in the biological visual cortex.

The Problem with Independence

For decades, the "holy grail" of natural image statistics has been Independent Component Analysis (ICA). The primary goal of ICA is to find a linear transformation that makes the components of an image as statistically independent as possible.

While ICA successfully produces "edge filters"—which are remarkably similar to the simple cells in our brains—it has a major flaw. The resulting components aren't actually independent. Strong dependencies remain, and ICA only offers marginal quantitative gains over simpler methods like Principal Component Analysis (PCA).

This led the authors to ask a provocative question: Instead of struggling to minimize dependency, what happens if we try to maximize a simple form of it? Specifically, they explored tree-dependencies, where each component is allowed to depend on exactly one "parent" component.

The Tree-Dependent Solution

To capture these relationships, the authors propose a model that learns three elements simultaneously from image patches:

  1. A Filter Matrix ($W$): An orthonormal matrix that transforms image pixels into filter responses.
  2. A Tree Structure ($T$): A graphical model where each node is a filter response, and edges represent dependencies.
  3. Density Functions: A model for how connected nodes relate to each other, using Gaussian Scale Mixtures (GSM) to capture the "heavy-tailed" nature of image data.
Comparison of dependency structures

As shown above, this approach moves away from assuming components are entirely independent (a) or only dependent in isolated pairs (b), and instead connects them in a comprehensive tree structure (c). To learn this tree, the authors use the Chow-Liu algorithm, which finds the maximum spanning tree by calculating the mutual information between all pairs of filter responses. They then use gradient ascent to update the filters, maximizing the overall log-likelihood of the data. This iterative process allows the filters and the tree structure to co-evolve based on the statistics of natural scenes.

The Emergence of Complex Cells

By allowing for tree-dependencies, the model doesn't just find edges; it discovers how edges relate to one another. In the learned tree, neighboring filters tend to have very similar orientations, positions, and frequencies, but different phases.

This leads to a fascinating biological connection. The model’s likelihood function involves summing the squared outputs of parent-child filter pairs. This mathematically mirrors the "energy model" of complex cells in the primary visual cortex (V1). These biological cells are known to be selective for orientation but invariant to the exact phase (or position) of an edge.

Key Results and Performance

The tree model provides a significantly better fit for natural image data than previous methods because it embraces the inherent correlations between different edge detectors.

  • Higher Likelihood: Evaluated on a standard dataset of 50,000 natural image patches, the tree-based model achieved a mean log-likelihood of -144.8, outperforming both ICA (-157.0) and Independent Subspace Analysis (ISA, -159.4).
  • Biological Realism: Approximately 60% of the learned filter pairs met the standard physiological criteria to be classified as "complex cells."
  • Qualitative Superiority: When sampling new images from the model, the tree-dependent samples showed more realistic, long-range structures (like continuous contours) compared to the "speckly" noise typically produced by ICA samples.

Limitations and Future Directions

While highly successful, the model does have limitations. It assumes the filter matrix $W$ is orthonormal to simplify the math, whereas real biological systems likely use overcomplete, non-orthogonal representations. Additionally, learning the tree structure and filters iteratively is computationally intensive.

Looking forward, the authors suggest this model could be highly effective for practical applications like image denoising. By using belief propagation on the learned tree structure, it may be possible to remove noise while carefully preserving the intricate dependencies that define natural textures and edges.

Summary updated

Abstract

We propose a new model for natural image statistics. Instead of minimizing dependency between components of natural images, we maximize a simple form of dependency in the form of tree-dependency. By learning filters and tree structures which are best suited for natural images we observe that the resulting filters are edge filters, similar to the famous ICA on natural images results. Calculating the likelihood of the model requires estimating the squared output of pairs of filters connected in the tree. We observe that after learning, these pairs of filters are predominantly of similar orientations but different phases, so their joint energy resembles models of complex cells.
BibTeX
@inproceedings{zoran2009tree,
  author = {Daniel Zoran and Yair Weiss},
  title = {{The 'tree-dependent components' of natural scenes are edge filters}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2009},
  url = {https://openreview.net/forum?id=r1bp2uWuZS},
}
Figure