Inhibition benefits neural system identification.

Area of Science:

• Computational Neuroscience
• Deep Learning
• Systems Neuroscience

Background:

• Neural system identification uses empirical data to model neuron stimulus-response functions.
• Deep neural networks enhance predictive performance but often omit crucial neural features like inhibition.
• Inhibitory interactions are vital for nonlinear neural computation in visual systems.

Purpose of the Study:

• To incorporate inhibition as an inductive bias into deep models for neural prediction.
• To investigate the influence of subtractive and divisive inhibition on learned neural transfer functions.
• To assess the impact of these operations on biologically plausible representations and specific visual computations.

Main Methods:

• Developed deep neural networks incorporating difference-of-Gaussian (subtraction) and within-channel divisive normalization (division) for neural prediction.
• Trained models on visual response data to learn stimulus-response functions.
• Performed in silico experiments to analyze learned kernels and the effects of inhibition on surround suppression and cross-orientation inhibition.

Main Results:

• Incorporating subtractive or divisive operations maintained predictive performance and yielded biologically plausible kernels.
• Both operations benefited the learning of surround suppression.
• Subtraction and division differentially affected activation and weight sparsity, and neither improved cross-orientation inhibition learning.

Conclusions:

• Inhibitory mechanisms, modeled via subtraction or division, can be effectively integrated into deep neural networks for improved neural prediction.
• These models offer insights into how inhibition shapes neural computations, particularly surround suppression.
• The findings highlight distinct effects of subtractive and divisive inhibition on neural representations and computations.