Interpretability of Multivariate Brain Maps in Linear Brain Decoding: Definition, and Heuristic Quantification in Multivariate Analysis of MEG Time-Locked Effects

Abstract

Brain decoding is a popular multivariate approach for hypothesis testing in neuroimaging. Linear classifiers are widely employed in the brain decoding paradigm to discriminate among experimental conditions. Then, the derived linear weights are visualized in the form of multivariate brain maps to further study spatio-temporal patterns of underlying neural activities. It is well known that the brain maps derived from weights of linear classifiers are hard to interpret because of high correlations between predictors, low signal to noise ratios, and the high dimensionality of neuroimaging data. Therefore, improving the interpretability of brain decoding approaches is of primary interest in many neuroimaging studies. Despite extensive studies of this type, at present, there is no formal definition for interpretability of multivariate brain maps. As a consequence, there is no quantitative measure for evaluating the interpretability of different brain decoding methods. In this paper, first, we present a theoretical definition of interpretability in brain decoding; we show that the interpretability of multivariate brain maps can be decomposed into their reproducibility and representativeness. Second, as an application of the proposed definition, we exemplify a heuristic for approximating the interpretability in multivariate analysis of evoked magnetoencephalography (MEG) responses. Third, we propose to combine the approximated interpretability and the generalization performance of the brain decoding into a new multi-objective criterion for model selection. Our results, for the simulated and real MEG data, show that optimizing the hyper-parameters of the regularized linear classifier based on the proposed criterion results in more informative multivariate brain maps. More importantly, the presented definition provides the theoretical background for quantitative evaluation of interpretability, and hence, facilitates the development of more effective brain decoding algorithms in the future.

Predicting anxiety from wholebrain activity patterns to emotional faces in young adults: a machine learning approach

“…Note that contributions of individual regions were very small, however, suggesting that predictions were based on the overall pattern rather than on a small combination of regions. As noted in previous studies (Schrouff and Mourao-Miranda, 2018, Haufe et al, 2014, Weichwald et al, 2015, Kia et al, 2017), this however does not mean that the brain activity associated with anxiety (measured by the STAI-T score) is distributed over the wholebrain. For example, Schrouff and Mourao-Miranda (2018) have shown that when the predictive patterns are subtle (i.e.…”

Neural Decoding of Multi-Modal Imagery Behavior Focusing on Temporal Complexity

“…These shortcomings might be partly attributable to individual differences (see Section 4.2), but other influences must also be considered. Although we have visualized the trained models with the highest accuracy, accuracy and interpretability do not always correlate (94,95). The present study has demonstrated the improvement of decoding accuracy achieved by using complexity measure as an input, but when considering its interpretation, it might be better to apply some constraint to the interpretability, rather than merely improving the accuracy.…”

Interpreting weight maps in terms of cognitive or clinical neuroscience: nonsense?

“…This is an important conclusion as the univariate contrast is often referred to as the 'ground truth' and used for model optimization e.g. based on weight stability [3]. Regarding the machine learning models, FP are present for both machines, but only in cases of low SNR.…”

“…The latter shows the relative contribution of the individual features to the model and has been heavily used in the neuroimaging community to infer conclusions about brain structure/function. There has however been a recent debate on whether weight maps can provide information about the neural signals leading to a significant classification/regression model [1]- [3]. The authors of [1] indeed suggest that weight maps provide a poor recovery of the input neural signal and lead to false positives.…”