Erlend Aune took some time to talk about machine learning techniques at an event at Bakken & Bæck’s Oslo office last night. Specifically, he outlined what steps you can take when your data set appears to be too limited to achieve what you want to do.
Key topics from the talk:
- Transfer learning
- Augmenting image data by generating variations of a single sample by cropping, rotating, style transfer
- Sophisticated concept translation (CycleGAN, DAGAN)
- Synonym replacement to augment text data
- Active learning
- Meta learning
- Expected gradient lengths
PDF slides: Small data – Musings on what to do when training data is limited
Thanks to everybody who showed up, and to Bakken & Bæck for hosting the event!
