Putting Humpty together again: Decoding brain networks during real-world processing

点击翻译按钮获取中文摘要

An important task for cognitive neuroscience is to understand how the brain supports broad functions that are engaged in everyday life, like interpersonal communication. Great progress has been made by decomposing these functions into discrete processes that can be associated with activity in particular brain regions and networks. What we do not know, however, is if and how these processes operate together. Thus, our overarching goal will be to understand how brain networks supports real-world behaviours under more ecological conditions. PhD project: aims and description (limit 300 words) Participants will watch films while undergoing functional magnetic resonance imaging (fMRI). The general aim is decompose the resulting data into networks and to label functions of those networks using annotation and machine learning approaches. We then characterize if, when, and how different networks interact and change with experience and test novel hypotheses about how the brain supports real-world processing.YR2. Neuroimage collection (supervised by Skipper). We will use fMRI to acquire anatomical and functional brain images from people watching three uncut fils in their native language. Each movie will be viewed by five adults who have not previously seen the film. We will use standard preprocessing routines on the data.YRS2-3. Film annotation and decoding (supervised by Skipper and Griffin). Features in films should be labelled in detail to decode the brain processes associated with those features. Hand annotation of features for even one film requires an enormous amount of time. To reduce this time, we will, first, use existing text-based descriptions and automated labelling to provide detailed time locked features. Second, we will use the brain data itself to identify film times for acquiring more detailed annotations through crowd-sourced approaches. The result of both is very rich text based descriptions of the films that are ready for machine learning decoding. This will be done by dividing data into training, validation, and test datasets and using the annotations to test how effectively networks can be labelled with different algorithms. Machine-learning approaches that might be most successful (support vector machines, convolutional neural networks, random forests) produce systems whose operations are non-trivial to understand. Thus, a component of the work will involve development of methods to probe solutions, to discover the spatial/temporal activity patterns used for predictions. These serve as the foundation for understanding how the brain processes natural stimuli.