Regularized Spectral Spike Response Model: A Neuron Model for Robust Parameter Reduction.

Area of Science:

• Computational Neuroscience
• Biophysics

Background:

• Current biological neuron models face challenges with hyperparameter optimization and overparameterization, restricting their use in biologically realistic simulations.
• These limitations hinder the development of accurate and efficient computational models of neuronal function.

Purpose of the Study:

• To introduce a novel neuron model, the Regularized Spectral Spike Response Model (RSSRM), designed to overcome the limitations of existing models.
• To demonstrate the efficiency and accuracy of RSSRM in simulating biological neuron behavior.

Main Methods:

• Developed the Regularized Spectral Spike Response Model (RSSRM) by incorporating regularization techniques to constrain parameter count.
• Implemented a fitting strategy that avoids explicit hyperparameter selection.
• Conducted twenty firing simulation experiments to evaluate model performance.

Main Results:

• RSSRM demonstrates superior performance compared to other neuron models, even after significant parameter pruning (>99%).
• With only 100 parameters, RSSRM achieved a 25% better RMSE in membrane potential prediction and a 99% better VRD.
• Achieved a 55% better F1-score and 24% better spike timing accuracy (±1.4 ms) than average models with the same parameter count.
• RSSRM exhibits high computational efficiency, with 10 KB memory usage and 1 ms inference runtime, outperforming the Izhikevich model.

Conclusions:

• The Regularized Spectral Spike Response Model (RSSRM) offers a highly accurate and computationally efficient alternative for biological neuron modeling.
• RSSRM's design effectively addresses overparameterization and the need for hyperparameter optimization, making it suitable for diverse biologically realistic tasks.