Contrast normalization contributes to a biologically-plausible model of receptive-field development in primary visual...

Area of Science:

• Neuroscience
• Computational Neuroscience
• Visual Processing

Background:

• Primary visual cortex (V1) neurons exhibit contrast normalization, where responses to stimuli decrease with stimulus complexity.
• This normalization is often attributed to inhibitory connections between V1 neurons.
• The Bienenstock, Cooper and Munro (BCM) learning rule models V1 neuron tuning but produces redundant population codes.

Purpose of the Study:

• To investigate if incorporating contrast normalization into the BCM rule can generate a more biologically realistic model of V1.
• To determine if this modified rule can lead to sparse, overcomplete representations of visual information.
• To explore the role of contrast normalization in the developmental plasticity of V1 receptive fields.

Main Methods:

• Modified the Bienenstock, Cooper and Munro (BCM) neural network learning rule by integrating a contrast normalization mechanism.
• Trained the modified BCM rule on natural images to observe the resulting neural representations.
• Analyzed the properties of the learned receptive fields and population coding efficiency.

Main Results:

• The BCM rule augmented with contrast normalization learned an efficient, sparse, and overcomplete representation of visual input.
• This modified rule produced model neurons with stimulus selectivity that better matches real V1 neurons.
• The learned representations suggest that neurons respond to distinct features of visual stimuli.

Conclusions:

• Contrast normalization, when incorporated into Hebbian learning rules like BCM, can lead to efficient and sparse neural representations.
• This suggests a developmental role for contrast normalization in shaping V1 receptive fields for diverse feature detection.
• The findings provide a more accurate computational model for understanding visual information processing in the mammalian V1.