3 ms·
In complete sincerity, I think that speeding up Turing-complete probabilistic programming to the kinds of inference speed we can get in the gradient-descent tra
by eli_gottlieb 9y ago
In complete sincerity, I think that speeding up Turing-complete probabilistic programming to the kinds of inference speed we can get in the gradient-descent training of deep neural networks would be a "change the world"-level advance for ML/AI.
- RJTrolo 9y agoWe already have that - variational inference based algorithms like BBVI use gradient descent for training
- eli_gottlieb 9y agoVariational inference also only works for continuous probability models, so it can't be used for most interesting use-cases of probabilistic programming.