Meta-Learning approaches in Deep Neural Networks

 

Antreas Antoniou

Meta-learning in deep neural networks is a research area that has recently shown much potential in learning better models that can be trained to have various desirable properties. Some very notable ones include converging in fewer steps, learning new optimizers, learning new architectures, learning weights that allow fast adaptation on new tasks, learning to learn models that allow high generalization on one-shot learning schemes or learning a small model that can generate the weights of a larger model that performs well on a given dataset and architecture and generalizes well to new datasets and even larger architectures (deeper architectures).

Supervisors: Amos Storkey & Tim Hospedales

 

 

Previous: Identifying Structure in Multitask Learning                         Next: Data-Efficient Deep Learning in Theory and Practice