Building sensitive models of cognition using interpretable Deep Learning

目前从脑成像数据预测行为和认知特征的方法受到限制，因为需要应用皮质组织的简化模型来促进数据集之间的比较。另一方面，深度学习允许直接比较成像数据，而无需使用此类先前的模型。因此，它有可能显著提高神经成像研究的灵敏度，就像它对自然图像分析所做的那样。不幸的是，深层网络进行预测的机制还没有被很好地理解，这使得人们很难知道数据的哪些特征对每项预测任务最具区分性。本项目的目标是开发用于可视化和解释深度学习算法的新技术，以：设计从脑成像数据中学习具有生物意义的表示的深度网络，并利用这些特征来提高对认知和行为机制的理解。本项目的具体目标是开发用于图形卷积深度学习算法的可视化和解释的新技术，以：1)设计从多通道皮质成像数据中学习具有生物意义的表示的深度网络。2)利用这些特征提高对早产儿和自闭症高危婴儿认知障碍机制的理解。

Current approaches for predicting behavioural and cognitive traits from brain imaging data are limited by a need to apply simplified models of cortical organisation to facilitate comparison between datasets. Deep Learning, on the other hand, allows direct comparisons of imaging data without the use of such prior models. It thus has the potential to significantly improve the sensitivity of neuroimaging studies as it has done for natural image analysis. Unfortunately, the mechanisms by which deep networks make predictions are not well understood making it challenging to know what features of the data are most discriminative for each predictive task. The goal of this project is to develop novel techniques for visualisation and interpretation of Deep Learning algorithms to:design Deep Networks that learn biologically meaningful representations from brain imaging data, and use these features to improve understanding of the mechanisms of cognition and behaviour.The specific objectives of this project are to develop novel techniques for visualisation and interpretation of graph convolutional Deep Learning algorithms to:1) Design Deep Networks that learn biologically meaningful representations from multi-modality cortical imaging data. 2) Use these features to improve understanding of the mechanisms of cognitive impairment in preterm born infants and babies at risk of Autism.