Towards conceptual generalization in the embedding space

(This is a whitepaper published in the early days of Kagi AI research)

A neural network in a self-driving car may properly react in most situations based on billions of images it has seen. But if there is a tree dangerously leaning over the road and is about to fall, the car will likely happily ignore it because it has no physical understanding of the world.

And in order to answer complex questions, we need a deeper understanding of the physical world that we live in.

In this whitepaper we get us one step closer to marrying physics and deep learning by using embeddings to represent physical states.

Towards conceptual generalization in the embedding space on Arxiv