A self-organizing cortical layer 4 model applied to large-scale imaging data analysis

Abstract

Event Abstract Back to Event A self-organizing cortical layer 4 model applied to large-scale imaging data analysis Simon Benjaminsson1* and Anders Lansner1 1 Royal Institute of Technology , Computational Biology, Sweden To find useful information in vast amounts of data is what our brains are particularly good at. It is also the ambition of many data analysis applications. Cortical models aimed at capturing the information processing capabilities of the brain are endowed with qualities that are suitable for large-scale data analysis, such as scalability, parallelization and an organization driven by the data statistics through learning. We have previously suggested a cortical layer 4 model as part of an attractor associative memory system [1]. It allows for transforming raw sensory data into a sparse and distributed code that reflects statistical properties of the input. More specifically, self-organization of the input sensors is facilitated by data dependencies. This leads to their grouping in a hypercolumnar structure. Together with a feature extraction step it underlies a hierarchical and modular cortical architecture. As well as using sensory data as input, any type of information can be utilized. We demonstrate this on resting-state fMRI data, employing the cortical layer 4 model to explore the statistics in a large multi-subject dataset. The voxels are treated as input sensors and the resulting groupings reflect the resting-state networks. The capability of the model to rapidly form different types of hypercolumnar organizations without excessive computational cost is exploited in this data analysis application to rapidly visualize resting-state networks along with their subnetworks.