3 ms·
The nice thing that saves ML from being only another way for AFGAM to dominate and force all the innovation to a highly concentrated group of geniuses is transf
by narrator 9y ago
The nice thing that saves ML from being only another way for AFGAM to dominate and force all the innovation to a highly concentrated group of geniuses is transfer learning! You can download Imagenet trained brains and get them to classify stuff just as good as in the competitions and then adapt them to various different applications. Google has been nice and releasing some of these models and they'll form the basis of open robot brains, etc. in the future. There will probably be a Open Robot distro pre-packaged with a bunch of these brains and applications in the future.
Little companies can adapt these brains to niches that are too small for the big guys.
- Eridrus 9y agoTransfer learning is great, but is still pretty limited; the main success stories are retraining CNNs from ImageNet to other datasets and word2vec/GloVE. In the space of all the things you would want to do with ML, this is pretty limited. [EDIT]: I stumbled upon a transfer learning for NER paper a second ago: https://arxiv.org/abs/1705.06273v1 https://arxiv.org/abs/1705.06273v1 And the results are not conclusive. You can can improve your results in truly data impoverished regimes, but you get diminishing returns as you add more data, so the only way to catch up is to get more data. So, you may be able to get reasonable performance with fewer samples, but you will still get better performance with more samples, and the more samples you get, the less benefit you get from transfer learning. This may not always hold, and I wonder if active learning paired with transfer learning would do better, but at the moment it's not convincing.