A Guided Tutorial on Modelling Human Event-Related Potentials with Recurrent Neural Networks.

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Machine Learning Applications in Neuroscience

Background:

• Computational models are crucial for generating hypotheses about event-related potentials (ERPs) in cognitive neuroscience.
• Cognitive neuroscientists may encounter technical hurdles in implementing these complex computational models.
• Recurrent Neural Networks (RNNs) offer a powerful framework for modeling dynamic neural processes like ERPs.

Purpose of the Study:

• To provide a tutorial on developing RNN models for ERP waveforms, making them more accessible to cognitive neuroscientists.
• To demonstrate the application of RNNs using the P3 component evoked by visual stimuli.
• To facilitate the use of computational modeling for understanding ERP generation mechanisms.

Main Methods:

• Development of RNN models using supervised learning with experimental event representations and ERP labels.
• Optimization of the RNN by minimizing mean-squared-error loss to approximate grand-average ERP waveforms.
• Analysis of RNN model behavior, including classification of hidden units and principal component analysis (PCA) of temporal responses.

Main Results:

• The RNN successfully approximated the grand-average ERP waveform when trained to link input representations with multiple ERP labels.
• The study demonstrates a method for evaluating RNN behavior as a model of computational principles underlying ERP generation.
• Techniques for analyzing RNN internal dynamics, such as hidden unit classification and PCA, are presented.

Conclusions:

• This tutorial successfully demonstrates a practical approach for implementing RNNs in ERP research.
• The presented methods enable cognitive neuroscientists to build and analyze computational models of ERPs.
• The approach facilitates hypothesis generation and provides insights into the computational basis of neural responses.