Simplify Building Two-Tower Matching Encoders to Improve Nearest-Neighbor Recommendation Results

The two-tower encoder is an approach to develop embeddings with supervised matches, as opposed to unsupervised single encoders. It was popularized in a paper by YouTube for recommendations. We'll demonstrate how you can easily train an embedding on Google's Cloud AI Platform with prebuilt algorithms. We'll show how you can configure different encoders like BERT and Swivel into the two-tower approach based on your use case. Lastly, you'll see how leveraging powerful GPUs can greatly speed up training time.