Hey, in this issue: how machine learning improves Google search, first many to many machine translation system (with code), learning with practically no data, detecting deepfakes with 80% accuracy, image-captioning AI better than human, machine-learning-based fast COVID-19 test, and more.

• Google details how it’s using machine learning to improve search (blog)

• Facebook released the first many-to-many multilingual model that translates between 100 languages without relying on English data (code, paper)

• AdaBelief optimizer: fast as Adam, generalizes as good as SGD, and sufficiently stable to train GANs (paper, code, video)

• Deconstructing Maxine, Nvidia’s AI-powered video-conferencing technology

• A radical new technique lets AI learn with practically no data

• When do we trust AI’s recommendations more than people’s?

• A new neural network can detect deepfakes at about 80% accuracy (for now)

• Nemo: data discovery at Facebook

• Microsoft’s image-captioning AI is describing pictures better than a human

• New Oxford machine learning-based COVID-19 test can provide results in under 5 minutes

Enjoy the newsletter? Help us make it bigger and better by sharing with your colleagues and friends.