Machine Learning Roundup 10/11/2016
Machine learning (ML) and deep learning (DL) content from the past 24 hours.
Note to readers: Big Analytics will rebrand as ML/DL soon.
Another great read from Nicole Hemsoth, this time on new developments in memory.
Adrian Colyer summarizes a paper from the Google DeepMind team on a framework for deep reinforcement learning with asynchronous gradient descent algorithms. The Google team achieved superlinear scalability with Deep-Q-Network (DQN) reinforcement learning when they tested asynchronous optimization methods.
If you’ve kept up with your assigned reading, you’re familiar with asynchronous machine learning from this paper linked last week and this article on DQN from last March. If you haven’t kept up with your assignments, how do you expect to compete in the AI economy?
Tesla Partners with NVIDIA
Fred Lambert reveals that Tesla will soon introduce Tesla Vision, a computer vision system built with NVIDIA’s CUDA GPU-powered parallel computing platform. Seeking Alpha reports. Merrill Lynch says “Buy.”
Kaggle announces a recruiting competition for Allstate, starting today and ending 11:59 pm, Monday 12 December 2016 UTC. No prize money, but Allstate will contact top-scoring entrants about possible job opportunities. No guarantees.
ML and Industry 4.0
Ben Hope describes the role of machine learning in Industry 4.0 (I4.0), the current buzzword for digitally transformed manufacturing.
Three Challenges for AI
Writing in The Verge, James Vincent describes three problems facing AI:
- Data-hunger. Machine learning systems need a lot more data than humans need to learn something useful.
- Overspecialization. We can build systems that recognize cat videos or play Atari, but we haven’t built systems that can do both.
- Opaqueness. If humans are to trust AI, systems must explain how they reach their conclusions.
With all the talk and buzz about Big Data, it’s a little surprising to hear people complain that we don’t have enough data.
Health and Medicine
Chuck Dinerstein, a physician, describes the increasing availability of algorithms and AI in medicine. He notes that questions of legal liability are as yet unresolved: what happens if an algorithm misdiagnoses a patient, who will be held accountable? He also wonders about adoption — will patients trust a black-box diagnosis? Finally, he asks how medical AI investors will recover costs and monetize the algorithms.
In DeviceTalk, a blog for the medical device industry, Ajit Singh describes key trends behind growing use of machine learning in clinical diagnostics.
Methods and Techniques
In ZDNet, Liam Tung describes the work of researchers at Google Brain who seek to develop a methodology that will reduce or eliminate unlawful bias in prediction algorithms.
Andrej Karpathy explains why Recurrent Neural Networks work well.
In MIT Technology Review, Will Knight profiles MobileEye, a startup that markets vision recognition technology with embedded deep learning. MobileEye and Tesla parted ways recently after a Tesla operating on AutoPilot hit a truck.
Google equips Chrysler minivans with self-driving systems.
In WallStreetDaily, David Dittman reports on the race to provide cars with AI.
BMW, Toyota, and Allianz invest in Nauto, a startup that markets technology that turns a fleet of vehicles into a self-learning network. Nauto’s technology package combines an in-vehicle camera, sensor hardware, a smart network and a continuously learning cloud to deliver a full, contextual view of vehicle and driver behavior.
Software and Services
Health Tech startup MYnd Analytics receives approval from the U.S. Centers for Medicare and Medicaid Services to serve as an Independent Diagnostic Testing Facility (IDTF). MYnd provides data and analysis services to clinicians.