A Comparison Between Pattern-Based Classifiers and Inverted Encoding Models for fMRI Decoding
Pattern-based classification has become the standard approach for the analysis of functional magnetic resonance imaging (fMRI) data. These classifiers use any discriminating information, that is present in the activity patterns across voxels, in order to decode stimuli or experimental conditions from the measured brain activity. Recently, however, explicit models of neural representation have become popular as an alternative approach to analysing fMRI data. These so-called encoding models describe explicitly how stimuli modulate the activity of neuronal populations, that is then eventually measured with fMRI. By inverting these models, they can also be used for decoding (i.e., for going from brain activity to stimuli). In this study, we systematically compare these two approaches commonly used for decoding. To do so, we use a linear support vector machine (SVM) and an inverted encoding model (IEM) to decode orientation from the activity patterns in primary and secondary visual cortex as measured with fMRI. Specifically, we investigate whether the two methods can produce comparable outputs and results. Additionally, we test whether these results are similarly influenced by different design and preprocessing parameters. We find that similar decoding accuracies and confusion matrices can be obtained from both methods. Moreover, the SVM and IEM decoding accuracies are affected in the same way by the different parameters: Decoding accuracies increase with more data used to train the classifier or used to fit the encoding model. For both methods, accuracies also first increase when more voxels are selected until reaching a plateau (SVM) or maximum (IEM). Furthermore, small amounts of spatial smoothing (3 or 4 mm) are beneficial for both approaches, whereas decoding accuracies decrease for larger amounts of smoothing. While both methods are comparable in terms of decoding accuracy, only IEM can be used to obtain channel response profiles that characterize the responses of neuronal subpopulations tuned to different feature values. Moreover, these channel response profiles are surprisingly unaffected by even large amounts of spatial smoothing, revealing a dissociation between the IEM response profiles and the IEM decoding accuracies.
Faculteit der Sociale Wetenschappen