Read about pinecone.io. Cut off a neural network that does image classification at the second to last layer and you have a system for creating meaningful vectorial representations. Then if you apply some distance metric to a bunch of these it turns out you can retrieve images that look similar to some query image.
Problem is, scaling and doing this in real-time is hard.
Use cases abound. This was only one. There are a lot of others too. Pinecone, Vespa.AI, Faiss…
Previously this type of tech was only available for the MAANGs of the world. Hope that helps
