Deep Representational Similarity Learning for Analyzing Neural Signatures in Task-based fMRI Dataset

Area of Science:

• Neuroscience
• Machine Learning
• Cognitive Science

Background:

• Representational Similarity Analysis (RSA) is vital for fMRI studies, measuring neural signature similarities across cognitive states.
• Existing RSA methods face limitations with high-dimensional, multi-subject fMRI data and are restricted by linear transformations or fixed nonlinear kernels.

Purpose of the Study:

• To introduce Deep Representational Similarity Learning (DRSL), a deep extension of RSA designed for complex fMRI datasets.
• To overcome the limitations of traditional RSA by employing flexible, subject-specific nonlinear transformations.

Main Methods:

• DRSL utilizes a multi-layer neural network to map neural responses into a linear space, enabling customized nonlinear transformations per subject.
• Gradient-based optimization is employed for efficient analysis of large datasets by processing data in batches.

Main Results:

• DRSL demonstrates superior performance compared to state-of-the-art RSA algorithms on multi-subject fMRI datasets.
• The method effectively analyzes similarities in neural patterns for diverse cognitive tasks, including visual stimuli, decision making, flavor perception, and working memory.

Conclusions:

• DRSL offers a powerful and flexible deep learning framework for advanced representational similarity analysis in neuroimaging.
• The proposed method significantly improves the analysis of high-dimensional fMRI data, paving the way for deeper insights into cognitive processes.