Convolutional neural network models applied to neuronal responses in macaque V1 reveal limited nonlinear processing

Area of Science:

• Computational neuroscience
• Computer vision
• Neuroscience

Background:

• Early models proposed primary visual cortex (V1) neurons act like Gabor filters with simple nonlinearities.
• Recent convolutional neural network (CNN) studies suggest V1 involves more complex nonlinear computations than previously thought.

Purpose of the Study:

• To investigate whether receptive field size, rather than inherent complexity, explains the performance of lower VGG-19 layers in predicting V1 responses.
• To compare the predictive power of VGG-19 and AlexNet models for macaque V1 neural activity.

Main Methods:

• Compared VGG-19 and AlexNet CNNs for predicting macaque V1 responses to natural and synthetic images.
• Analyzed the effect of receptive field size and image size on model performance.
• Utilized a Gabor pyramid model to assess nonlinear contributions like normalization and contrast saturation.

Main Results:

• AlexNet's early layers, with larger receptive fields, better predicted V1 responses than VGG-19's lower layers.
• A modified AlexNet matched VGG-19's performance with fewer nonlinear computations.
• Reducing input image size shifted VGG-19's optimal layer, supporting the receptive field size hypothesis.

Conclusions:

• V1 feedforward responses can be explained by a limited number of nonlinear processing stages.
• Receptive field size is a critical factor in CNN models predicting V1 activity.
• The findings challenge the notion of extensive nonlinear computations in early visual processing.