Convolutional neural network models describe the encoding subspace of local circuits in auditory cortex.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Convolutional neural networks (CNNs) are powerful tools for modeling neural sensory encoding.
• The complexity of CNNs hinders understanding the specific computations driving their performance.
• Interpreting deep learning models in neuroscience remains a significant challenge.

Purpose of the Study:

• To develop a method for identifying informative sensory dimensions within CNNs.
• To create a framework for interpreting deep-learning-based neural encoding models.
• To understand how CNNs represent auditory information in the brain.

Main Methods:

• Trained a CNN on single-neuron auditory cortex data from ferrets presented with natural sounds.
• Computed each neuron's linear tuning subspace using dimensionality reduction on CNN output gradients.
• Developed a linear-nonlinear subspace model to predict neural activity based on identified subspaces.

Main Results:

• The developed linear-nonlinear subspace model proved functionally equivalent to the original CNN.
• Analysis revealed that neural population responses sparsely tiled a shared stimulus subspace.
• Observed differences in encoding properties across cell types and CNN layers.

Conclusions:

• The study establishes a novel framework for interpreting deep-learning-based neural encoding models.
• The findings provide insights into how auditory cortex neurons represent complex natural sounds.
• This approach facilitates a deeper understanding of neural computations underlying sensory processing.