Alex Tamkin on Self-Supervised Learning and Large Language Models
The Gradient: Perspectives on AI
00:00
The Future of Neural Networks for Contrastive Learning
There's no scientific basis for say, it was just sort of people tried and these things worked. And I think that kind of got us interested in, well, is there some sort of better way that you might be able to learn these types of data augmentations for contrastive learning? We found this cool image to image neural network, which takes an image and outputs another image. But it wasn't clear initially, like, how should we train this network to produce good augmentations?"
Transcript
Play full episode
