Decoding fMRI Data: A Comparison Between Support Vector Machines and Deep Neural Networks

Abstract Multivoxel pattern analysis (MVPA) examines the differences in fMRI activation patterns associated with different cognitive conditions and provides information not possible with the conventional univariate analysis. Support vector machines (SVMs) are the predominant machine learning method in MVPA. SVMs are intuitive and easy to apply. The limitation is that it is a linear method and mainly suitable for analyzing data that are linearly separable. Convolutional neural networks (CNNs), a class of AI models originally developed for object recognition, are known to have the ability to approximate nonlinear relationships. CNNs are rapidly becoming an alternative to SVMs. The purpose of this study is to compare the two methods when they are applied to the same datasets. Two datasets were considered: (1) fMRI data collected from participants during a cued visual spatial attention task (the attention dataset) and (2) fMRI data collected from participants viewing natural images containing varying degrees of affective content (the emotion dataset). We found that (1) both SVM and CNN are able to achieve above chance level decoding accuracies for attention control and emotion processing in both the primary visual cortex and the whole brain with, (2) the CNN decoding accuracies are consistently higher than that of the SVM, (3) the SVM and CNN decoding accuracies are generally not correlated with each other, and (4) the heatmaps derived from SVM and CNN are not significantly overlapping. These results suggest that (1) there are both linearly separable features and nonlinearly separable features in fMRI data that distinguish cognitive conditions and (2) applying both SVM and CNN to the same data may yield a more comprehensive understanding of neuroimaging data. Key points We compared the performance and characteristics of SVM and CNN, two major methods in MVPA analysis of neuroimaging data, by applying them to the same two fMRI datasets. Both SVM and CNN achieved decoding accuracies above chance level for both datasets in the chosen ROIs and the CNN decoding accuracies were consistently higher than those of SVM. The heatmaps derived from SVM and CNN, which assess the contribution of voxels or brain regions to MVPA decoding performance, showed no significant overlap, providing evidence that the two methods depend on distinct brain activity patterns for decoding cognitive conditions.