Deep Learning | Learning with Small Data

[@arne8780]

- mitigate small data sets: use auxiliary tasks, domain knowledge
- auxiliary tasks: transfer learning, multi-task learning, meta learning
- transfer learning (domain adaptation): supervised vs. unsupervised
- 10:20 unsupervised transfer learning: no labeled data in target domain
- goal: similar feature space after embedding
- MMD (maximum mean discrepancy): distance of distributions
- 14:11 multi-task learning
- hard-parameter sharing
- soft-parameter sharing
- layer routing
- 16:55 meta-learning