123 - Robust NLP, with Robin Jia
NLP Highlights
00:00
Certifiably Robust Training for Neural Networks
The big idea there is that neural networks conveniently have this nice structure where they're very modular, right? So they're basically composed of a sequence of elementary operations and some computation graph. And so for each kind of node in the computation graph, we're going to say, if I know some constraints on what the possible input to this node is, I can give you some corresponding bounding box around what the possible outputs of this layer are. Can this approach be extended beyond looking at word vectors? Like could you also imagine doing this on a sentence level?
Transcript
Play full episode
