3 ms·
Has #1 really been proven to be false ? Transfer learning using the first few layers pertained on imagenet or a related task have consistently given 1-2% impro
by screye 8y ago
Has #1 really been proven to be false ?
Transfer learning using the first few layers pertained on imagenet or a related task have consistently given 1-2% improvements in scores... As recently as mid 2018.
This is especially for complex tasks like VQA
- gok 8y agoTransfer learning (or student-teacher training) isn't really the same thing as pre-training as it was being talked about 10 years ago. And in those days the claim wasn't that it helped a bit, but that it made training many deeper networks possible at all.