Decoding fMRI data with support vector machines and deep neural networks.

Area of Science:

• Neuroimaging
• Cognitive Neuroscience
• Machine Learning in Neuroscience

Background:

• Multivoxel pattern analysis (MVPA) uses fMRI data to study cognitive conditions.
• Support vector machines (SVMs) are common in MVPA but limited to linearly separable data.
• Convolutional neural networks (CNNs) can model nonlinear relationships, showing promise for fMRI analysis.

Purpose of the Study:

• To compare the performance of SVM and CNN models on fMRI datasets.
• To understand the similarities and differences between SVM and CNN in fMRI analysis.

Main Methods:

• Applied both SVM and CNN to two fMRI datasets: one from a visual spatial attention task and another from viewing affective images.
• Compared decoding accuracies and classification patterns between the two methods.

Main Results:

• Both SVM and CNN achieved above-chance decoding for attention and emotion processing.
• CNN consistently yielded higher decoding accuracies than SVM.
• SVM and CNN decoding accuracies were largely uncorrelated, with non-overlapping derived heatmaps.

Conclusions:

• SVM and CNN utilize different neural features for classification.
• Combining SVM and CNN approaches may provide a more comprehensive understanding of neuroimaging data.