4 ms·
This style transfer idea, though eye candy and sometimes impressive, seems to be the core application of deep learning these days. François Chollet tweeted some
by felippee 10y ago
This style transfer idea, though eye candy and sometimes impressive, seems to be the core application of deep learning these days. François Chollet tweeted something like that two years ago about the prisma app (yes it was 2015). He back then anticipated many other killer apps behind the corner but it seems not much had materialized. It's 2017 now and people are still super excited about yet another style transfer network. I'm not even sure where this would be practically useful, aside from being yet another photoshop/instagram filter. Am I the only one skeptical about this?
- rmellow 10y agoI understand your position, but it comes from a false assumption that this is just about images. This is a demo that uses images as a way to "wow" the audience. However, the kind of network used here (GAN) can be used in domains other than images (text, financial data). Imagine if you trained a network to generate fraudulent financial data and another to become an expert at catching fraud, each feeding back into the other's skill. This is the concept of GANs at heart and definitely disruptive if correctly executed.
- felippee 10y agoIt is not entirely obvious that such extrapolations would succeed. GAN is not a new idea (new in the context of deep learning) but automatic generation of adversarial samples is a known and exploited subject (even e.g. in the field of evolutionary strategies). So this is still the promise: look, it works on the images but there are all these wonderful applications possible. Its 2017 now, I'll check back in 2019 to see if anything has changed.
- hugozap 10y agoIt's 2017 and cars drive themselves. Isn't that exciting?
- aisofteng 10y agoThey don't. If you've ever chanced upon one of Uber's prototypes, the engineer in the driver's seat takes over often.