Diversity priors for learning early visual features.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Computer Vision

Background:

• Visual neurons in area V1 exhibit sparsity and selectivity.
• Previous computational models often focus on sparsity or selectivity independently.
• The diversity of receptive fields, crucial for these properties, is underexplored in learning.

Purpose of the Study:

• To investigate if diversity priors can lead to the discovery of biologically plausible early visual features.
• To demonstrate that diversity priors can simultaneously enforce sparsity and selectivity in neural network activations.
• To explore the role of receptive field diversity in computational models of early vision.

Main Methods:

• Utilized restricted Boltzmann machines (RBMs) to model natural image statistics and learn visual features.
• Introduced a novel diversity prior during the training of RBMs.
• Analyzed the learned receptive fields and neuron activation properties.

Main Results:

• The diversity prior successfully promoted simultaneous sparsity and selectivity of neuron activations.
• Learned receptive fields showed a high degree of similarity to biological data from area V1.
• The discovered visual features demonstrated strong generative capabilities in image reconstruction tasks.

Conclusions:

• Diversity priors are effective in learning biologically relevant early visual features using RBMs.
• Exploiting receptive field diversity offers a unified approach to achieving sparsity and selectivity.
• This method enhances the biological plausibility and generative power of computational vision models.