Amazon wants to automate the human picking process at its warehouses. This is hard, robots nead to pick up and place arbitrary objects from a potential pool of millions, so Amazon has been running a competition for three years on the subject. Smart way to motivate great teams to work on this.
This post describes an approach where a so-called "hyper-network" is used to generate the weights of another network. Thereby improving this network. Hyper-networks provide an abstraction that is similar to what is found in nature: the relationship between a genotype - the hyper-network - and a phenotype - the main network.
Great post by Sebastian Ruder, with best practices on many aspects of NLP: word embeddings, network construction, multi-task learning, attention and more.
Machine learning at Stripe has a foundation built on Python and the PyData stack, with a production system written in Scala. This talk covers the ML Infra team’s work to bridge the serialization and scoring gap between Python and the JVM, as well as how ML Engineers ship models to production.