>>12887862>limitsIt really "depends" on what you decide ML is
all the generative models/re-enforcement learning is taking off like a goddamn rocket, we haven't seen them applied to many fields at all. transformers (as tards are fighting about in this thread) are pretty nice in what they can do; I've replaced some of my RNN work with them, specifically the really long-term dependency sequences I use to predict.
Graph-based methods combined with something like skip-gram models as inputs to transformers/other models seem to be a new step, taking some really robust high-dimensional relationships and encoding them down
model-stacking is slowly being explored (AlphaStar was real goddamn fun to watch, I implement about 60% of the networks that make up AlphaStar in my work for different things in the drug space, super fun to watch it put together:
https://deepmind.com/blog/article/alphastar-mastering-real-time-strategy-game-starcraft-ii)
latentGANs/styleGANs are a neat idea. Stacking and freezing weights/increasing the complexity of training input allows some really nice control over the generated images (and not just images, but drug-design as well)
Basically, we're at the baby's first steps of getting ML implemented- and now we are finding new model architectures that generalize better and can do more (RNN's exploded onto the scene in what, 2015? read "the unreasonable effectiveness of RNNs" and now they are almost old-dust compared to what we have now).
People who take those google-intro to ML or follow the tensorflow/pytorch tutorials and build simple classification models and go "huh that's it for ML? ML IS DEAD! THERE IS NOTHING MORE" don't really know much about the field.
