2020 [arXiv]
Whats Unique One classic bottleneck of current deep learning systems is that it considers the dataset as the task and hence suffers from the ability to generalize. Thoughtfully crafted meta-learning system can be helpful here.
A good read on a possible meta-learning setup from Prof Manning's team, where K^N (where k is latent clusters, and N is a number of labels) datasets are created from the original one to better represent task-based learning, rather than dataset based learning.
How It Works
- Following figure gives overview of the approach
Source: Author
-
Interestingly, it validates its hypothesis on a artificially generated datasets of sinewave, and then apply it on the real NLI datasets.
-
It uses BERT to get embedding representations, and used K-means to cluster in K-clusters. After which K^N datasets are created for meta-learning setup.