Contextual Integration in Cortical and Convolutional Neural Networks.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Neurons are hypothesized to represent sensory input via probability distributions and perform probabilistic inference.
• Lateral connections in neural circuits exhibit non-random patterns and modulate responses to stimuli.
• Cortical connectivity mapping reveals cell-type specific connections, including excitatory and inhibitory neuron interactions.

Purpose of the Study:

• To develop a neuronal network model approximating Bayesian inference for feature probability estimation.
• To link observed neural connectivity patterns to computational principles.
• To investigate the role of lateral connections in enhancing robustness and performance of artificial neural networks.

Main Methods:

• Proposed a neuronal network model for Bayesian inference of feature presence probabilities.
• Analyzed lateral connection dependencies on activity correlations and global inhibition.
• Incorporated learned lateral connections into convolutional neural networks (CNNs) and evaluated performance on noisy datasets.

Main Results:

• The model's predicted connectivity qualitatively matched empirical data from mouse primary visual cortex.
• Neurons with similar orientation tuning showed stronger connectivity, with modest spatial extents for both excitatory and inhibitory connections.
• CNNs with learned lateral connections demonstrated improved robustness to noise and enhanced performance on noisy MNIST datasets.

Conclusions:

• Lateral connections are crucial for contextual integration and approximating Bayesian inference in neural circuits.
• The model provides a framework for understanding cell-type specific connectivity and its computational roles.
• Combining supervised and unsupervised learning through lateral connections offers a promising approach for real-world vision tasks, especially with unreliable inputs.