News in artificial intelligence and machine learning

From Jan 22nd thru 12th Feb. Prefer to receive this in your inbox? Sign up here.

Technology news, trends and opinions

• Go, Marvin Minsky, and the Chasm that AI Hasn’t Yet Crossed. The big news these past few days is around DeepMind’s AlphaGo system. Using a combination of supervised learning and reinforcement learning, the authors show that they’re able to defeat 99.8% of computer Go programs and a European (human) champion. Paper here.
• Often the bane of the life of software engineers, perhaps there’s hope in sight? Automatic bug-repair system fixes 10 times as many errors as its predecessors.
• An informative summary from Re.Work’s Deep Learning Summit held in SF. I attended the event and was pleasantly surprised by the trials run by image/vision deep learning companies (Bay Labs, MetaMind, Clarifai) in the healthcare space, specifically on diagnostic support systems. There’s such a compelling use case here that I hope said companies have streamlined access into clinics.
• Along these lines, IBM’s Automated Radiologist Can Read Images and Medical Records too.
• Artificial Intelligence Offers a Better Way to Diagnose Malaria. I’ve sat in front of a microscope for many hours in the past and I can definitely see value in applying deep learning systems for automated feature recognition. In fact, many researchers still use outdated tools like ImageJ to segment images and count features, manually :-/
• Google’s Self-Driving AI Counts as a “Driver,” According to the Feds. The plot thickens…
• Wired profiles Andy Rubin, founder of Android, and quite a fascinating character. His new gig, Playground, seeks to create, fund and support companies writing software and hardware building blocks for an AI-enabled future.

Research, development and resources

• Asynchronous Methods for Deep Reinforcement Learning, Google DeepMind and University of Montreal. Stabilising deep reinforcement learning often requires lots of memory and computation because it uses experience replay: the idea of storing and sampling from an agent’s historical experience conducting a task. Here, the DeepMind team show that asynchronously running multiple agents in parallel on multiple instances of the environment is instead more effective and less computationally intensive. Check out this video for a demo application of an agent navigating a 3D video game environment.
• A model explanation system, Northrop Grumman Corporation. Here, the author presents a general framework for explaining black box models, specifically linear classifiers.
• No more Autobahn! Scenic Route Generation Using Googles Street View, University of Bremen and Hasselt University. In a similar vein to Daniele Quercia’s Happy Maps project (watch his talk at our 2015 AI Summit), the authors present a vision-based system to generate scenic driving routes using Google Street View images.
• Deep Tracking: Seeing Beyond Seeing Using Recurrent Neural Networks, University of Oxford. The authors present an end-to-end RNN-based learning method for taking in a stream of raw sensor data (e.g. a scene where a robot sees lots of moving objects) and finding the positions of these objects even if they’re occluded from view and where the ground truth isn’t available. Video explanation here.
• Brain4Cars: Car That Knows Before You Do via Sensory-Fusion Deep Learning Architecture, Cornell University, Stanford University and Brain of Things Inc. This work explores the problem of anticipating a car driver’s next actions to safeguard said driver from eventual dangers that might result based on their current environment (e.g. turning and hitting an unseen bicycle). Using RNNs with LSTMs that act on videos, vehicle dynamics, GPS, and street maps, the system can anticipate manoeuvres 3.5 seconds before they occur in real time.

Venture capital financings and exits

$93m raised by 22 companies and $250m in exit value created, including the following notable transactions:

• DataRobot, which allows data scientists to build and implement predictive applications using open source and custom machine learning, raised $33m Series B led by NEA.
• Diffbot, a business building a universal database of structured information from the web, raised $10m led by Tencent.
• Sky, a leading European entertainment company, purchases a $10m stake in DataXu, a marketing and analytics company for the adtech business. The company has now raised $65.6m since 2007.
• Trifacta, which helps data analysts process raw data into clean and structured formats for downstream work, raised a $35m Series D from Accel, Greylock and others.
• SwiftKey, the maker of the popular predictive swipey keyboard for Android and iOS used by 300m people, was acquired by Microsoft for $250m. Congrats, team! Great news for the market.

Anything else catch your eye? Just hit reply! I’m actively looking for entrepreneurs building companies that build/use AI to rethink the way we live and work.