Modern Machine Learning as a Benchmark for Fitting Neural Responses.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning in Neuroscience

Background:

• Generalized linear models (GLMs) are common for predicting neural activity but may not capture all explainable variance.
• Assessing the performance of different encoding models is crucial for understanding neural computations.

Purpose of the Study:

• To compare the predictive performance of machine learning methods against traditional GLMs for neural encoding.
• To evaluate the effectiveness of feedforward neural networks, XGBoost, and stacked ensembles in predicting neuronal firing rates.

Main Methods:

• Predicted spike counts in macaque motor (M1) and somatosensory (S1) cortices using reaching kinematics.
• Predicted rat hippocampal cell activity based on open field location and orientation.
• Compared GLMs with feedforward neural networks, XGBoost, and stacked ensembles.

Main Results:

• XGBoost and stacked ensembles consistently yielded more accurate spike rate predictions than GLMs.
• Machine learning methods demonstrated reduced sensitivity to feature preprocessing.
• These advanced models effectively identified relationships between features and neural activity missed by simpler models.

Conclusions:

• Machine learning approaches, particularly XGBoost and ensembles, provide accurate neural encoding models.
• These methods serve as valuable benchmarks for evaluating simpler predictive models in neuroscience.
• Advanced machine learning enhances the ability to decode neural representations.