>>11366197Just get a PhD
>>11366185The latter.
"statistics & machine learning" doesn't exist so don't fall for the scam.
>>11366159CS PhDs in deep learning (or ML in general) are extra hot currently, but it's a bubble that won't last forever. Other CS PhDs not so much.
A PhD is required for any kind of science/research role while a master's will lock you into engineering roles. If you like to copy-paste last year's new meme paper that doesn't even work as well as previous methods just because it's popular and the bossman said to do it, go master's only. If you want to prove the equivalence between stable lstms and fully connected networks, and that unstable lstms can boost performance only for short-range relations, then dig into why these are equivalent, you need a PhD.
>>11366101Unsupervised is probably not going anywhere further. In the first place you can't really evaluate it in a vacuum like you're suggesting which is the main reason why there's no real reason to suspect it can improve by any kind of concerted research effort (i.e. without working only in domain-directed ways in which case it's not really unsupervised anymore).
RL however definitely has a lot of leeway, though it will need an explosive leap to improve, since its main problem is the asymptotic efficiency being no better than monte carlo-based search (it's more efficient only in the multiplicative constant, and only just, requiring intense reward function hand-design effort to get any results). While it's possible that RL is also bottomed out, I wouldn't buy that one. But the AI winter happened regardless of the potential of ML/DL, so that doesn't mean the bubble won't pop despite RL probably having a lot of space to grow.
WaveNet and GLOW are literally rebranded ancient shit everyone was doing a decade ago. GANs are cool and game theoretical methods might lead to good shit in the future, though no leads so far. Theory is catching up at last at least.
