ldn ml_1 gp

-most know what GP is, great

Rest in peace David

-can take ... output is

..

instead of doing linear thing, map data into feature space potentially very high-dim. then do linear model in high-dim feature space

-GP is a Bayesian kernel regression

What would he see?

-this is just background

-NNs are all the rage when undergrad -RNNs for parsing natural language -...didn't work as no data, slow computers, ideas for how to train RNNs not as sophisticated as they are now -all rage in early 80s and 90s -in that era Radford Neil looked at RNNs from Bayesian POV -...turned out LeCun had done this earlier

-most know what GP is, great

-prior doesn't have to be Gaussian

-a lot has happened since 80s and 90s -if you look at them, ideas are quite similar with no. of diff's

-architectural differences: dropouts,LSTMs,...but also vastly larger datasets -also have vastly larger compute resources:{GPU,cloud}s: -,with all of that has come media investment, -,and a lot of media hype.