Deep learning models for brain-behaviour associations

点击翻译按钮获取中文摘要

Deep learning approaches based on Canonical Correlation Analysis (CCA), such as Deep Canonical Correlation Analysis (DCCA, Andrew et al. 2013) and Deep Canonically Correlated Autoencoders (DCCAE, Wang et al. 2015), are machine learning techniques that can learn linear and non-linear transformations of two views of the same samples/subjects in a common/latent space, in which the learnt representations are highly correlated. This project will aim at employing these transformations to create mappings between multivariate brain patterns and behaviour/clinical patterns by using large samples of healthy individuals as well as individuals with mental health disorders. The developed generative models will then be used to test specific hypothesis of brain-behaviour associations related to mental health disorders. For example, given specific clinical/cognitive profiles, the model will be able to generate the corresponding brain patterns and vice-versa. This innovative modelling approach is expected to shed light on the underlying mechanisms of brain-behaviour association. Further applications of the developed framework will include utilising the latent space to perform patient stratification according to mental health disorders, and using the generative models to synthesize missing data to aid predictive studies.