Search

This thesis investigates online, offline and transfer learning for decision making tasks using a combination of behavioural experiments, computational modelling and Electroencephalography (EEG). Our experiments used a new set of decision-making tasks, in which the appropriate response depended on the linear or nonlinear combination of multiple stimulus features, and were developed to have better ecological validity than many previous tasks in the literature. The first study, in chapter 2, outlines the contextual settings in which representations of the environment can be learnt online. We manipulated the temporal structure of trials, and nature of stimulus-response mappings, and showed their effects on performance and declarative learning. We fitted a Latent Cause Model (LCM) of participants behaviour and derived measures that we used to gain insight into the representations formed. In chapter 3 we used EEG to identify the multiple successive stages of representation learning preceding decisions and following feedback. We used a Computational-EEG approach in which subject-specific LCM variables were used to predict a subject's EEG data, and found evidence of feature representation in sensory regions and more complex representations in frontal regions. In chapter 4 we shifted the focus to offline learning, by examining the effect of a period of quiet wakefulness on performance in the same task. We found that quiet wakefulness significantly improved the generalization of previously learnt associations. Finally, in the last study, in chapter 5, we investigated how knowledge acquired in one task can be transferred to another. We borrowed the concept of shared subspaces from the multitask learning literature and showed that this provides a useful framework for the study of human transfer learning. Taken as a whole, the thesis shows how humans form representations online and offline, and how extracted knowledge can be transferred to new tasks.

Medulloblastoma is a malignant embryonal tumour of the cerebellum which most commonly affects children. A subset of tumours is thought to arise from cerebellar granule cell precursors (GCPs) that fail to undergo normal neuronal development, following the hyper-activation of the Sonic hedgehog (Shh) signalling pathway. To identify candidate genes that might be important for medulloblastoma pathogenesis, I investigated patterns of gene expression and alternative splicing in 14 paediatric medulloblastoma and five normal cerebellar samples using Affymetrix Human Exon arrays. Statistical analysis of the gene expression data identified a group of medulloblastomas with a molecular signature of Shh pathway activation. These tumours showed higher expression levels of genes involved in spindle formation, cytokines, and cell cycle regulation. Further studies using an in vitro mouse GCP cell culture model, in which Shh is necessary for the maintenance of the progenitor state, showed that a selection of candidate genes was also over-expressed when GCPs were cultured in the presence of Shh, as compared to control cells, as well as in human medulloblastoma cell lines. Ongoing and future in vitro experiments will investigate the potential role of candidate genes in sustaining the growth of precursor and tumour cells. Exon-level analysis of gene expression showed that abnormal expression of different transcript variants is likely to occur in medulloblastoma. I selected several examples of differential exon usage and validated these using an independent set of normal and tumour specimens. Tumour-associated splicing alterations were highly consistent, enabling clear separation of normal and cancer samples and in some cases of different medulloblastoma molecular subgroups. Interestingly, Shh-treated GCPs recapitulated the splicing patterns observed in the tumour samples for six out of the eight genes analysed, suggesting that the preferential expression of specific transcript forms is regulated during normal cerebellar development. The possible relationship between inappropriate splicing and medulloblastoma pathogenesis will be the subject of future investigation.