107 - Multi-Modal Transformers, with Hao Tan and Mohit Bansal
NLP Highlights
00:00
Is There Room for a Latent Alignment Model?
A latent alignment model embraces the fact that, as you say, you're always going to get a very incomplete description of the image. And like, i tries to detect at what level of depth you're showing stuff. S also related to coriculum learning, where you could be going from like, simpler to longer sentences. We can also define the notion of like, specificity in images and captions,. Where they basicallead as a way to calculate how a specific or how a slightly set deepa the description is so yet w i could see this happening both as a sort of inpard feature, additional feature sinary, or as learning iletenor ising tesin
Transcript
Play full episode
