Convolutional neural network models of V1 responses to complex patterns.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Understanding how neurons in the primary visual cortex (V1) process complex visual information is crucial.
• Traditional models often struggle to capture the full complexity of V1 neuronal responses to intricate stimuli.

Purpose of the Study:

• To evaluate the efficacy of convolutional neural network (CNN) models for predicting V1 neuron activity in awake macaque monkeys.
• To identify the specific components of CNNs that contribute to their performance in V1 modeling.

Main Methods:

• Modeling V1 neurons using a large dataset of complex pattern stimuli.
• Comparing CNN performance against established models like Gabor-based and generalized linear models.
• Systematically dissecting CNN components (e.g., thresholding nonlinearity, convolution) and using transfer learning with deep CNNs.

Main Results:

• CNN models significantly outperformed all baseline models in predicting V1 neuronal responses.
• Thresholding nonlinearity and convolution were identified as critical components driving CNN success.
• Higher layers of deep CNNs, encoding more complex features, showed superior performance, aligning with previous findings on V1 neural code complexity.

Conclusions:

• CNNs offer a powerful framework for modeling V1 neurons, surpassing conventional approaches.
• Specific architectural elements of CNNs are essential for accurately capturing V1 neuronal computations.
• Deep CNNs provide insights into hierarchical processing within the visual cortex.