Fast and robust estimation of spectro-temporal receptive fields using stochastic approximations

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Signal Processing

Background:

• Receptive fields (RFs) are crucial for understanding sensory neuron signal preferences and neural coding.
• Estimating neuronal RFs is essential but computationally intensive, especially with complex models or high-dimensional data.
• Existing methods face challenges with computational load in large-scale neural recordings.

Purpose of the Study:

• To develop a computationally efficient optimization scheme for estimating neuronal receptive fields.
• To accelerate the analysis of neural responses for complex RF models.
• To enable simultaneous monitoring of RF properties for multiple neurons.

Main Methods:

• Proposed an optimization scheme based on stochastic approximations (SA).
• Applied stochastic gradient descent (SGD) algorithms to generalized linear models (GLM) and classification-based RF estimation.
• Utilized random data subsets for faster computation instead of the full dataset.

Main Results:

• Stochastic approximations yield robust spectro-temporal receptive field (STRF) estimates.
• Preserved predictive power and tuning properties of STRFs in auditory midbrain recordings.
• Achieved high correlation (0.93) with full solution estimates in <10% of the computation time.
• Developed an on-line algorithm for simultaneous STRF monitoring of >30 neurons.

Conclusions:

• The proposed SA-based approach is beneficial for large-scale neural recordings.
• Offers a more comprehensive characterization of neural tuning compared to standard tuning curves.
• Enhances computational efficiency in RF estimation for neuroscience research.