Luo, Xiaoliang; (2022) Goal-directed cross-system interactions in brain and deep learning networks. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Thesis.pdf - Accepted Version Download (10MB) | Preview

Abstract

Deep neural networks (DNN) have recently emerged as promising models for the mammalian ventral visual stream. However, how ventral stream adapts to various goal-directed influences and coordinates with higher-level brain regions during learning remain poorly understood. By incorporating top-down influences involving attentional cues, linguistic labels and novel category learning into DNN models, the thesis offers an explanation for how the tasks we do shape representations across levels in models and related brain regions including ventral visual stream, HPC and ventromedial prefrontal cortex (vmPFC) via a theoretical modelling approach. The thesis include three main contributions. In the first contribution, I developed a goal-directed attention mechanism which extends general-purpose DNN with the ability to reconfigure itself to better suit the current task goal, much like PFC modulates activity along the ventral stream. In the second contribution, I uncovered how linguistic labelling shapes semantic representation by amending existing DNN to both predict the meaning and the categorical label of an object. Supported by simulation results involving fine-grained and coarse-grained labels, I concluded that differences in label use, whether across languages or levels of expertise, manifest in differences in the semantic representations that support label discrimination. In the third contribution, I aimed to better understand cross-brain mechanisms in a novel learning task by combining insights on labelling and attention obtained from preceding efforts. Integrating DNN with a novel clustering model built off from SUSTAIN, the proposed account captures human category learning behaviour and the underlying neural mechanisms across multiple interacting brain areas involving HPC, vmPFC and the ventral visual stream. By extending models of the ventral stream to incorporate goal-directed cross-system coordination, I hope the thesis can inform understanding of the neurobiology supporting object recognition and category learning which in turn help us advance designs of deep learning models.

Download activity - last month

Export as XMLJSONCSV

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as JSONXMLCSV

Archive Staff Only

View Item