RESUMEN
Explainable AI aims to overcome the black-box nature of complex ML models like neural networks by generating explanations for their predictions. Explanations often take the form of a heatmap identifying input features (e.g. pixels) that are relevant to the model's decision. These explanations, however, entangle the potentially multiple factors that enter into the overall complex decision strategy. We propose to disentangle explanations by extracting at some intermediate layer of a neural network, subspaces that capture the multiple and distinct activation patterns (e.g. visual concepts) that are relevant to the prediction. To automatically extract these subspaces, we propose two new analyses, extending principles found in PCA or ICA to explanations. These novel analyses, which we call principal relevant component analysis (PRCA) and disentangled relevant subspace analysis (DRSA), maximize relevance instead of e.g. variance or kurtosis. This allows for a much stronger focus of the analysis on what the ML model actually uses for predicting, ignoring activations or concepts to which the model is invariant. Our approach is general enough to work alongside common attribution techniques such as Shapley Value, Integrated Gradients, or LRP. Our proposed methods show to be practically useful and compare favorably to the state of the art as demonstrated on benchmarks and three use cases.
RESUMEN
Lateralization is a fundamental characteristic of many behaviors and the organization of the brain, and atypical lateralization has been suggested to be linked to various brain-related disorders such as autism and schizophrenia. Right-handedness is one of the most prominent markers of human behavioural lateralization, yet its neurobiological basis remains to be determined. Here, we present a large-scale analysis of handedness, as measured by self-reported direction of hand preference, and its variability related to brain structural and functional organization in the UK Biobank (N = 36,024). A multivariate machine learning approach with multi-modalities of brain imaging data was adopted, to reveal how well brain imaging features could predict individual's handedness (i.e., right-handedness vs. non-right-handedness) and further identify the top brain signatures that contributed to the prediction. Overall, the results showed a good prediction performance, with an area under the receiver operating characteristic curve (AUROC) score of up to 0.72, driven largely by resting-state functional measures. Virtual lesion analysis and large-scale decoding analysis suggested that the brain networks with the highest importance in the prediction showed functional relevance to hand movement and several higher-level cognitive functions including language, arithmetic, and social interaction. Genetic analyses of contributions of common DNA polymorphisms to the imaging-derived handedness prediction score showed a significant heritability (h2=7.55%, p <0.001) that was similar to and slightly higher than that for the behavioural measure itself (h2=6.74%, p <0.001). The genetic correlation between the two was high (rg=0.71), suggesting that the imaging-derived score could be used as a surrogate in genetic studies where the behavioural measure is not available. This large-scale study using multimodal brain imaging and multivariate machine learning has shed new light on the neural correlates of human handedness.
