2 ms·
Indeed, although I'm not as familiar with transfers this far (that is, fine-tuning models is really common with image recognition tasks, but often you take a mo
by joshuamorton 10y ago
Indeed, although I'm not as familiar with transfers this far (that is, fine-tuning models is really common with image recognition tasks, but often you take a model trained to recognize objects in images, and train it to recognize different objects in images, but you're still recognizing objects in images).
This feels a bit different, in that yesterday I would have had no strong intuition that "a char-rnn can detect text sentiment better than sota". Looking now, I can rationalize that idea. I get why it might make sense, but it was non-obvious.
Do you disagree with any of that? (and if you do, I'd love to see this distance of transfer in literature, its always cool to read up on these things)