BMW Machine Learning Weekly — Week 3
March 1 — March 7, 2018
News about Machine Learning (ML), Artificial Intelligence (AI) and related research areas.
ML Complicates Effects of new EU Data Privacy Regulations
New EU Regulations on Personal Data, the General Data Protection Regulation (GDPR), which come into effect on 25 May 2018, may be difficult to combine with ML techniques. The key theme of the regulation is that everyone owns their own data. Any company must therefore explicitly request permission to use any personal data, explaining why it would like to do so, and for how long. Later, this permission can be withdrawn at any time. The worry is that it can become problematic to explain how an ordinary consumer’s personal data might be used to train algorithms to infer outcomes for others. Even if the consumer then consents, the consumer can withraw this permission at any time, which technically may require an ML algorithm to “unlearn” (and thereby forget) how an individual’s specific data adapted the learning of the algorithm.
ML in the US Military
Google has partnered with the United States Department of Defense to help the agency develop AI algorithms to analyzing drone footage using TensorFlow. According to the Pentagon, the initial goal of project “Maven” is to provide the military with advanced computer vision, enabling the automated detection and identification of objects in as many as 38 categories captured by a drone’s full-motion camera. Maven provides the department with the ability to track individuals as they come and go from different locations. The collaboration set off a firestorm among employees of the technology giant, because the project raised important ethical questions about the development and use of ML.
When Bananas Look Like Toasters
“Adversarial examples” (or, AIs on hallucinogens) are increasingly becoming a problem. Adversarial examples are specially designed optical illusions fooling computers into thinking they recognize something for an object that it is not. Examples range from mistaking a picture of a rifle for a helicopter to mistaking a banana for a toaster after a sticker was put on the banana that confuses computer vision systems.
Researches — including Ian Goodfellow, the creator of generative adversarial networks (GANs) — published a paper in which they revealed that they have been able to generate first photos that fool both humans and computer vision algorithms. The research into finding altered images that cannot fool the human mind are important for real-life use cases such as autonomous driving. It is necessary, that a system can see every stop sign, no matter how much it might be altered.
Noteworthy
- Morphing Cars Into Helicopters
Geneva Motor Show: a collaboration between Italdesign, Airbus and Audi presented a two-seater Smart Car-sized monocoque, which can ride along on a base of wheels as a regular car, or — with the help of a humongous drone module — be hoisted into the air for vertical flight. Read more… - Bringing AI to the Airport
The tech company SITA revealed plans to let AI take over the management of baggage at airports. The aim is to minimize the amount of missing and mishandled bags. Similarly, Unisys software isbringing ML algorithms to border control. - Goodbye Green Screen, Hello AI!
Google is replacing its green screen with an IA tool by bringing pro-level video-editing techniques to its phones. Currently, the tool is in limited beta mode. Read more… - Explore Historical Life Photos
Using Ai to sort millions of historical Life photos, Google published a nice tool to create a new, searchable archive of Life photographs. Read more…
