10 July 2019
Embeddings are everywhere in machine learning yet usually not very visible. We use embedding layers when one hot encodings don’t cut it or to visualize relations between words but usually we don't appreciate the crucial role they play in artificial intelligence.
This talk tells the story of embeddings by showing not only how they are used in different deep learning models, but also how they can really help us get an intuition about how these models function.
We'll start by looking at MNIST handwritten digits, before moving on to embedding words for fun and profit. After that we'll look at modern image recognition systems and how that recognition is just a nearest neighbor search on the produced embeddings.
We'll close by looking at how embeddings are used in machine translation and how that work opens up the tantalizing perspectives of using thought vectors towards general intelligence.