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

Area of Science:

• Neuroscience
• Computational Auditory Neuroscience
• Machine Learning in Neuroscience

Background:

• The auditory cortex processes complex spectro-temporal sound features using nonlinear combinations.
• Convolutional neural networks (CNNs) show promise as accurate encoding models for neural activity evoked by natural sounds.
• Interpreting the computational mechanisms underlying CNN performance in neuroscience remains challenging due to their complexity.

Purpose of the Study:

• To develop a method for visualizing the tuning subspace captured by CNNs in the auditory cortex.
• To understand the computational properties enabling CNNs' superior performance in predicting neural responses.
• To link deep learning models with established neuroscience concepts for better interpretability.

Main Methods:

• Recorded single-unit data from the primary auditory cortex (A1) of ferrets using high-density microelectrode arrays.
• Fit a CNN to predict neural activity evoked by a large natural sound set.
• Measured the dynamic spectro-temporal receptive field (dSTRF) of the CNN and used principal component analysis (PCA) to define a low-dimensional tuning subspace.

Main Results:

• A subspace model, derived from PCA of the CNN's dSTRF, predicted neural activity nearly as accurately as the full CNN.
• Visualization in the tuning subspace revealed diverse nonlinear responses, including contrast gain control and phase invariance.
• Neurons within local populations tiled the tuning subspace, forming a sparse representation; inhibitory neurons exhibited distinct tuning patterns.

Conclusions:

• The developed subspace visualization method provides a framework for interpreting deep learning models in neuroscience.
• CNNs capture complex, nonlinear spectro-temporal processing in the auditory cortex, which can be effectively represented in a low-dimensional subspace.
• This approach reveals insights into neural coding principles, such as sparse coding and the functional roles of different neuron types in A1.