Meta-learning

The deep learning revolution created created a massive change in the performance and perception of AI technologies. However, deep learning techniques require large amounts of hand-labeled data which limits applications in certain domains where such amount of data is not available.

Meta-learning is an example of learning-to-learn approaches. The idea is to meta-train an entity such as a neural network-based classifier or robot policy on a set of tasks that are related to, but not identical to the the target task of interest. Meta-training prepares the representations and gradients of the network to learn the target with a minimum amount of data (one or few examples).

Our work in this direction develops techniques to make the meta-learning phase unsupervised, reducing the need for labeled data with several orders of magnitudes. We are also working towards applying meta-learning to the circumstances of mobile computing.

Publications

• S. Khodadadeh, L. Bölöni, and M. Shah. Unsupervised Meta-Learning For Few-Shot Image Classification. In Proc. of Thirty-third Conference on Neural Information Processing Systems (NeurIPS-2019), pp. 10132–10142, December 2019.   BibTeX   Download