Graph Neural Networks are Finally Useful—Here's Why #airevolution #ai #datascience

Añadido: 2026-05-17

[00:00:00]This is the So, this great question. So, my group pioneered graph neural networks.

[00:00:07]This is the next generation of that.

[00:00:08]Okay. And the beautiful thing is that it's it's general. It's not about graphs. It's about databases. So, everyone has a database.

[00:00:18]>> yeah. Cuz the problem with graph neural networks has always been no one has a graph.

[00:00:23]Nobody has a graph. But now I'm saying any database is a graph. You can now train what we call relational graph transformers directly on the raw data and you get all [clears throat] the benefits of, let's say, true deep learning. Right? Think of computer vision pre-deep learning, post-deep learning. Post-deep learning. Pre-deep learning, it was about detecting edges and sift features and Gabor filters and then training a classifier.

[00:00:48]Post-deep learning era, learn directly on the raw pixels of the image.

[00:00:52]We are doing the same thing, right?

[00:00:55]Where we say, let's learn directly on your database.

[00:00:58]How do we do that because, you know, a database is not an image. Database is not a sequence of text. We say you represent the database as a graph and now you can your neurons can directly reason over that graph. And you get all the benefits that we have already seen, you know, in in computer vision as well as in natural language understanding with large language models who just, again, learn over a sequence.