Sequential learning in the form of shaping as a source of cognitive flexibility.
Doctoral thesis, UCL (University College London).
Humans and animals have the ability to quickly learn new tasks, a rapidity that is unlikely to be manageable by pure trial and error learning on each task separately. Instead, key to this rapid adaptability appears to be the ability to integrate skills and knowledge obtained from previous tasks. This is assumed for example in the sequential build-up of curricula in education, and has been employed in training animals for behavioural experiments at least since the initial work on shaping by Skinner in 1938. Despite its importance to natural learning, from a computational neuroscience point of view the question of sequential learning of tasks has largely been ignored. Instead, learning algorithms have often been devised that are capable of learning from an initial naive state. However, it is known that simply training sequentially with the same algorithms can often harm learning through interference, rather than enhance it. In this thesis, we explore the e�ffects of sequential training in the form of shaping in the cognitive domain. We consider abstract, yet neurally inspired, learning models and propose extensions and requirements to ensure that shaping is bene�ficial. We take the 12-AX task, a hierarchical working memory task with rich structure, de�fine a shaping sequence to break the hierarchical structure of the task into separate smaller and simpler tasks, and compare performance between learning the task in one fell swoop to that of learning it with the help of shaping. Using a Long Short-Term Memory (LSTM) network model, we show that learning times can be reduced substantially through shaping. Furthermore, other metrics such as forms of abstraction and generalisation may also show differential e�ffects. Crucial to this, though, is the ability to prevent interference, which we achieve through an architectural extension in the form of "resource allocation". Finally, we present initial, human behavioural data on the 12-AX task, showing that humans can learn it in a single session. Nevertheless, the task is sufficiently challenging to reveal interesting behavioural structure. This supports its use as a candidate to probe computational aspects of cognitive learning, including shaping. Furthermore, our data show that the shaping protocol used in the modelling studies can also improve averaged asymptotic performance in humans. Overall, we show the importance of taking sequential task learning into account, provided there is additional architectural support. We propose and demonstrate candidates for this support.
|Title:||Sequential learning in the form of shaping as a source of cognitive flexibility|
|Open access status:||An open access version is available from UCL Discovery|
|UCL classification:||UCL > School of Life and Medical Sciences > Faculty of Life Sciences > Gatsby Computational Neuroscience Unit|
Archive Staff Only