Efficient Visual Coding: From Retina To V2

Honghao Shan, Garrison CottrellICLR 2014 · arXiv · OpenReview

This paper presents a hierarchical computational model that simulates the early stages of the human visual pathway. By applying principles of efficient coding—specifically Sparse PCA and Independent Component Analysis (ICA)—the authors demonstrate how the brain might automatically learn to represent visual information, from the simple center-surround filters of the retina to the complex contour-detecting neurons of the V2 cortex.

TL;DR

The researchers developed a multi-layer hierarchical model that uses Sparse PCA and ICA to mimic the early human visual system. The model successfully replicates the receptive field properties of neurons from the retina through to area V2 and suggests that certain V2 neurons, previously dismissed by experimentalists as "noise," actually play a vital role in detecting object contours.

What Problem Does This Paper Address?

The human visual system is organized hierarchically (Retina $\rightarrow$ LGN $\rightarrow$ V1 $\rightarrow$ V2), yet most computational models have struggled to explain this entire pathway within a single, consistent framework. While "efficient coding theory"—the idea that the brain evolves to reduce redundancy in sensory input—has explained V1 edge detectors well, it hasn't been as successfully applied to the retina or the higher-order features of V2.

The authors aim to bridge this gap by creating a model that learns universal visual features from natural scenes without supervision. They also seek to provide a theoretical basis for neurophysiological observations that have puzzled experimentalists, such as why some V2 neurons don't seem to exhibit clear orientation tuning.

Methods

The authors propose a variant of the Recursive ICA (RICA) model. The architecture is modular, with each "layer" of the visual system consisting of three main steps:

Hierarchical model architecture showing the sequence of PCA, ICA, and nonlinear activations.
  1. Sparse PCA (sPCA): This reduces the dimensionality of the input while imposing a "sparsity" constraint. This mimics the biological bottleneck of the retina, where 100 million photoreceptors must be compressed into just 1 million ganglion cell responses.
  2. Sparse Coding / ICA: This expands the dimensionality again to capture higher-order statistical structures. This step mimics the expansion of signals in the visual cortex (V1).
  3. Coordinate-wise Nonlinearity: The outputs are processed through a nonlinearity (taking absolute values and "Gaussianizing" the distribution) to "format" the data for the next layer in the hierarchy.

This process is repeated: the output of the first layer (Retina/V1) becomes the input for the second layer (V1 complex cells/V2).

Datasets

The model was trained and evaluated using the Kyoto Natural Scene Dataset, consisting of 62 high-resolution color images of natural environments. Images were preprocessed by converting them to grayscale, down-sampling, and normalizing to zero-mean and unit-variance. The model was trained on 16x16 image patches for the first layer and 32x32 patches for the second layer.

Key Insights and Results

  • Layer 1 (Retina/LGN & V1 Simple): Unlike standard PCA, Sparse PCA naturally produces "center-surround" receptive fields. This suggests that the circular "on-center/off-surround" filters found in the retina are a direct result of the need to compress information efficiently. The subsequent ICA layer successfully produced oriented, local edge detectors.
Receptive fields learned by the model: center-surround filters from sPCA (left) and oriented edge detectors from ICA (right).
  • Automatic Pooling (V1 Complex): When the model applies sPCA to the output of the first ICA layer, it automatically performs "pooling." It groups together oriented edges of the same orientation in a local area—exactly what V1 complex cells do in the brain.
  • Layer 2 (V2): The model's second ICA layer produced features matching V2 cells, including those with uniform orientation tuning across their field and those with non-uniform tuning (like "corner" detectors). The model's distribution of cell types closely matched the proportions found in macaque monkey V2 studies.
  • Correcting the Literature: A famous study by Van Essen's group discarded about 13% of recorded V2 neurons because they didn't show clear orientation tuning. This paper argues those neurons weren't "junk"—they are a predicted part of an efficient coding system and likely serve to capture object contours.

Limitations and Future Work

  • Feed-forward Only: The current model is strictly feed-forward. In the actual human brain, there is significant "top-down" feedback from higher visual areas to lower ones, which this model does not yet account for.
  • Static Images: The primary results are based on static grayscale images. A more robust exploration of temporal dynamics (motion) and chromaticity would be a logical next step.
  • Experimental Validation: The authors' prediction regarding the "discarded" V2 neurons provides a clear, testable hypothesis for future neurophysiological experiments.

Summary updated

Abstract

The human visual system has a hierarchical structure consisting of layers of processing, such as the retina, V1, V2, etc. Understanding the functional roles of these visual processing layers would help to integrate the psychophysiological and neurophysiological models into a consistent theory of human vision, and would also provide insights to computer vision research. One classical theory of the early visual pathway hypothesizes that it serves to capture the statistical structure of the visual inputs by efficiently coding the visual information in its outputs. Until recently, most computational models following this theory have focused upon explaining the receptive field properties of one or two visual layers. Recent work in deep networks has eliminated this concern, however, there is till the retinal layer to consider. Here we improve on a previously-described hierarchical model Recursive ICA (RICA) [1] which starts with PCA, followed by a layer of sparse coding or ICA, followed by a component-wise nonlinearity derived from considerations of the variable distributions expected by ICA. This process is then repeated. In this work, we improve on this model by using a new version of sparse PCA (sPCA), which results in biologically-plausible receptive fields for both the sPCA and ICA/sparse coding. When applied to natural image patches, our model learns visual features exhibiting the receptive field properties of retinal ganglion cells/lateral geniculate nucleus (LGN) cells, V1 simple cells, V1 complex cells, and V2 cells. Our work provides predictions for experimental neuroscience studies. For example, our result suggests that a previous neurophysiological study improperly discarded some of their recorded neurons; we predict that their discarded neurons capture the shape contour of objects.
BibTeX
@inproceedings{shan2014efficient,
  author = {Honghao Shan and Garrison Cottrell},
  title = {{Efficient Visual Coding: From Retina To V2}},
  booktitle = {International Conference on Learning Representations},
  year = {2014},
  url = {https://openreview.net/forum?id=CRge-EDLedRUr},
}
Figure