Retaining skills under distribution shifts: sequential Bayesian inference, reinforcement learning and applications

Thesis

Modern machine learning models, such as neural networks which are the focus of this thesis, have been shown to be extremely powerful tools for learning function mappings from inputs to outputs, for example, for image-to-categories or audio-to-text tasks. However, when learning over a set of sequential tasks, one after another, they struggle to recall past tasks. Neural networks, forget previous abilities they have learned when learning new tasks. The requirement for neural network...

Expand abstract

Authors: Kessler, SC

• DOI: 10.5287/ora-bx6vmn55x
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English