Related Experiment Videos
A comparison of natural-image-based models of simple-cell coding
B Willmore1, P A Watters, D J Tolhurst
1Department of Physiology, University of Cambridge, Downing Street, Cambridge CB2 3EG, UK. bw200@cam.ac.uk
Perception
|January 6, 2001
Summary
The Olshausen-Field (O&F) model and principal components analysis (PCA) are compared for visual coding efficiency. After accounting for preprocessing and filter size, the O&F model shows slightly sparser coding than PCA, requiring dispersal measures for clear distinction.
Area of Science:
- Computational neuroscience
- Visual processing
- Information theory
Background:
- Modeling simple-cell receptive fields in V1 is crucial for understanding visual coding.
- The Olshausen-Field (O&F) model and principal components analysis (PCA) are prominent approaches.
- Both models aim to link natural image statistics to neural coding principles.
Purpose of the Study:
- To evaluate the sparseness and dispersal of O&F and PCA models for efficient visual coding.
- To clarify ambiguous definitions of sparseness and dispersal in the literature.
- To investigate the impact of image preprocessing on model performance.
Main Methods:
- Formulated specific definitions for sparseness and dispersal for model comparison.
- Analyzed the influence of preprocessing steps like spectral pseudo-whitening and logarithmic transformation.
- Assessed the effect of filter size and normalization on sparseness and dispersal measures.
Main Results:
- Preprocessing significantly affects both sparseness and dispersal measures.
- PCA filter sparseness is dependent on receptive field size.
- Normalization of filter means and variances impacts dispersal measures.
Conclusions:
- The O&F model yields slightly sparser visual codes than PCA when controlled for preprocessing and filter size.
- The difference in sparseness between O&F and PCA is less pronounced than anticipated.
- Dispersal measures are essential for effectively differentiating between O&F and PCA coding strategies.