Last week Yoshua Bengio made the announcement that MILA would end development on Theano, bringing to an end a deep learning project that had been a mainstay in open source Deep Learning for over a decade.

COMMENTS

Things have changed quickly: new open source projects like TensorFlow and PyTorch have significantly disrupted the Deep Learning landscape for python developers, and front end tools like Keras continue to open new options for accessing industry projects like DL4J with python. Deep Learning is now a standard tool in the Machine Learning scientist’s toolkit, where just a few short years ago tinkering in Theano to test out strictly R&D ideas was the closest most came to applying Deep learning in industry. While Bengio’s announcement may have felt inevitable (though possibly sooner than expected), it marks the end of an era and a milestone in the new age of Deep Learning for industry.

John Giannandrea, who leads AI at Google, drew attention to the threat of systematic bias learned by AI systems. Suggesting that this risk elevates the importance of questioning “black box” systems for Machine Learning driven operations.

COMMENTS

Giannandrea’s comments dovetail with the formation of a new AI ethics research unit at DeepMind (owned by google). Further, his insight against trusting “black box” systems is far from trivial. Most of the Deep Learning advances that have enabled the AI industry significantly are notoriously opaque. Examples highlighted in the TechCrunch article “The racist, fascist, xenophobic, misogynistic, intelligent machine", punctuate the fact that the ML bias threat is real and should be taken seriously by Machine Learning scientists and business operators alike.

NIH provided one of the largest chest x-ray dataset to scientific community. The goal is to provide Deep Learning researchers the training data that may potentially enable reliable Machine Learning based screening in the future.

DeepMind announced this week that after recent advances in speed efficiency, WaveNet algorithms would be replacing the voice of Google Assistant. WaveNet is a convolutional neural net based algorithm, used to generate raw audio waveforms. According to published A/B tests, the resulting voices are capable of producing better and more realistic-sounding speech than existing techniques.

This glossary is a comprehensive resource for any Machine Learning scientist. No doubt the link will become an essential bookmark for both novices and experts.

This paper identifies RNN based language models used to automate the generation of fake online reviews for products and services. Not only are these attacks cheap and therefore more scalable, but they can control rate of content output enabling them to evade detection through anomalously rapid crowdsource methods. Survey-based results show these reviews not only evade human detection, but also score high on “usefulness” metrics by users.

Trevor Hastie’s books have long been a staple in any Machine Learning scientist’s library. This text is equal parts, practitioners reference guide and education tool. A solid technical tour of core concepts and techniques.