>>13271918I work in ML in drug discovery. There's some real hot shit in the ML space right now that everyone is exploring currently. My list is:
>1 Meta-learning/k-shot learning / reinforcement learningBig topics I bundled together. Motivated by the fact that human only need a few examples to generalize well, while current ML needs huge datasets. Started with siamese networks, now prototypical networks (stupid easy) and meta-agnostic meta learning are at the front. re-enforcement learning has taken some leaps with AlphaZero and the offspring, we're seeing something resembling general learning for the first time, imo.
>attentionWhen attention mechanisms were introduced, it was the first time in my career I felt we could actually call something in ML "ai-like". This are most "last 5 years progress" and less "Future progress", but we're seeing some great adaptations and uses, and we'll see ripples in advancements across the sciences as the mechanism get more widely adopted and altered.
>graph convolutional networksAlso on the cusp of something big. The ability to generalize graph networks in the same way we do convolutional networks would be a goddamn game changer. We're dancing around this right now; Kipfs work is a teaser of what's to come. Most of us are pretty sure we'll have a *true* graph convolutional network in 5 years or so. I've been playing with graph-based networks and I'm surprised at how good they are at generalizing to unknown targets/data for things like link prediction.
As for other stuff, uhh
>crispr treatmentsno-brainer
>neural regenerationreal interesting topic. Even though I'm in ML, my background is molecular biology; there has been some jaw-dropping progress in spine regeneration, particularly with some finding that if you knockdown some transcription factors you get reversion back to early neurons (growth cone and all) that can then restore spinal connections in mice with severed spines. I imagine that's going somewhere fast.
