Noise Correlations for Faster and More Robust Learning.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Distributed population codes are common in the brain, presenting a challenge for downstream neurons to learn appropriate readouts.
• Stimulus-independent noise correlations are frequently observed in neural populations, despite theoretical implications for reduced representational capacity.

Purpose of the Study:

• To investigate whether stimulus-independent noise correlations can simplify the learning of neural readouts by constraining learning to task-relevant dimensions.
• To determine how manipulating noise correlations affects the speed and robustness of learning in neural networks performing perceptual tasks.

Main Methods:

• Utilized neural network models to simulate a perceptual discrimination task.
• Manipulated noise correlations among similarly tuned units independently of the population signal-to-noise ratio.
• Assessed the impact of noise correlations on readout learning speed and homogeneity of neuronal weights.

Main Results:

• Higher noise correlations among similarly tuned units led to faster and more robust learning.
• Noise correlations favored homogenous weight assignments to neurons within functionally similar pools, potentially emerging via Hebbian learning.
• In multi-discrimination tasks, correlations across relevant features accelerated learning, while those across irrelevant features hindered it.

Conclusions:

• Noise correlations, when maintained without significant signal-to-noise ratio degradation, can enhance readout learning speed by constraining it to appropriate dimensions.
• These findings suggest a functional role for noise correlations in neural computation, aiding learning processes in both simple and complex tasks.