Google Uses Neural Networks to Translate Without Transcribing

New Scientist, 4/4/2017

Researchers at Google Brain are developing a method to translate speech in one language to text in a different language, using neural networks. They hope to improve over more conventional automatic methods, where the speech is transcribed into written text, and then the written text is translated. The older method can be cumbersome, and initial experiments show that direct speech-to-text translation seems less subject to error. The new method could especially help speakers of rare languages communicate with others around the globe.

AncestryAI Algorithm Traces Your Family Tree Back More Than 300 Years

Aalto University, 3/16/2017

An Aalto University graduate student is developing a new family-tree algorithm called AncestryAI that uses Finnish parish registers to generate a probable family tree. Register data comes from HisKi, an open genealogical database supported by volunteer data entry. AncestryAI users can leave feedback on the accuracy of their tree, which is used to further train the algorithm.

Quantifying Urban Revitalization

MIT News, 10/24/2016

Researchers have developed a system that estimates how safe visitors will perceive an area to be based on photographs of the area. The researchers began with a crowdsourcing effort to build an initial database of images and safety ratings of the area in the image. Over 1.4 million responses have then been used to train a machine-learning algorithm to identify these aspects automatically.