News in artificial intelligence and machine learning you should know about
From 13th April thru 3rd May. Referred by a friend? Sign up here. Want to share? Give it a tweet.
*I joined our friends at MEX on a podcast to explore artificial intelligence, user behaviour and how to design better digital experiences.
*Part 2 of my podcast with Nick Moran on investing in AI is now live.
Technology news, trends and opinions
Health-related AI 🔬
Just two weeks after Mass General Hospital announced their Clinical Data Science Center, news emerges that Imaging Advantage has partnered with MIT/Harvard researchers to deliver a deep learning-based product to help radiologists read X-Rays. The company has a reported 7 billion images in its database and is used in 450 radiology facilities in the US.
On the other end of the spectrum, remote patient monitoring (e.g. for chronically ill patients) is another frontier where machine learning can thrive. This piece argues that the use of biosensors and wearables make it possible to build models that learn to detect and alert patients who are at risk of deteriorating. I really do believe that we should shift our healthcare system from being reactive to preventative by investing more in real-time monitoring of physiology to drive early detection.
To make all of this happen, however, technology companies will need to have access to patient health records. Google DeepMind Health is in the media limelight again regarding its work with the NHS. New Scientist got hold of the data-sharing agreement between the company and the Royal Free NHS Trust in question. 1.6 million patients pass through its three London hospitals each year.
AI and society 🌎
Is the general population aware of the extent to which the algorithms that drive AI products affect us every day? This piece explores the notion of algorithmic accountability and implores those reporting on AI to peel back the veils that cloak the inner workings of algorithms.
Further on this topic, Rory Cellan Jones at the BBC runs a great podcast on the topic of Google’s dominance in search — the prime example of algorithms ruling our digital experiences. Do you think ‘Google’ is synonymous with ‘search’? *Read to the bottom to find out!
Nick Bostrom, author of Superintelligence, published a new working paper entitled Strategic Implications of Openness in AI Development in which he considers the objectives and most desirable forms of openness. In the short-term, openness expedites the dissemination of existing technologies. In the long-term, the consequences of openness (and cumulative existential risk) depend on whether we privilege currently existing people over potential future generations.
What the big incumbents are up to 🏢
In the past, software companies had to solve their own problems with storage, networking and compute before launching their products. Today, these challenges have been productised, abstracted and slotted into a company's P&L as costs of goods sold. This piece in Fortune explores Facebook’s desire to solve AI to the same degree. The company’s Head of Applied Machine Learning, Joaquin Candela, puts it this way: “We need to make AI be so completely part of our engineering fabric that you take it for granted.”
Google too have outlined how key AI is to all of their work. In this year’s Founders’ Letter to shareholders, Sundar Pichai shares these gems:
Google Photos, which launched less than a year ago, currently boasts 100 million monthly active users!
What’s next after mobile (note that Android now has 1.4 billion 30-day-active devices)? Sundar writes, “the next big step will be for the very concept of the “device” to fade away. Over time, the computer itself — whatever its form factor — will be an intelligent assistant helping you through your day. We will move from mobile first to an AI first world”.
Enterprise productivity, supercharged with AI, is next on the list. “We see huge opportunities to dramatically improve how people work. Your phone should proactively bring up the right documents, schedule and map your meetings, let people know if you are late, suggest responses to messages, handle your payments and expenses, etc.”
Google DeepMind’s staff roster has grown 2x since its acquisition 2 years ago, sourcing 65% of them from academia (namely UCL and Oxford).
Welcoming the up and comers! 🙌🏻
FT featured a piece on our newest investment, Numerai, exploring how the company uses homomorphic encryption to obfuscate complexity in financial markets data to open up the prediction problem to a global community of data scientists and machine learning enthusiasts. I do believe this approach of crowdsourcing problem solving, which has been demonstrated by Kaggle, can be vertically integrated to create a new breed of companies like Numerai.
MIT Tech Review run a story on Magic Pony Technology, the 12-strong London-based startup developing machine learning-based approaches for visual processing. Given how our serious video binging habits, mobile networks will start feeling the crunch, and this team might be their saving grace. Well done, chaps!
Research, development and resources
We’re in the lead up to several conferences, CVPR (computer vision), ICML (machine learning) and ICCC (computational creativity) so lots of papers are out!
Artistic Style Transfer for Videos, University of Freiburg. Last year, Matthias Bethge’s research group (*he will be speaking at Playfair AI 2016) demonstrated that convolutional neural networks (CNN) can be used to learn representations of artistic style from one painting such that it can be applied to real world photographs. In this work, Thomas Brox’s group builds on this work to show that style can be learned from a single image and transferred to an entire video sequence. Two improvements are made: a) To ensure that style consistency extends over longer video sequences when certain regions might be temporarily occluded, the authors use long term motion estimates; b) A multi-pass algorithm processes the video several times and alternates in the forward and backward direction to remove artefacts at image boundaries. Do check out the video results here. As an aside, there’s another neat paper on automatically colorising greyscale images using CNNs.
CVPR2016: Learning Deep Feature Representations with Domain Guided Dropout for Person Re-identification, The Chinese University of Hong Kong. This work explores the problem of training a neural network using datasets from multiple domains (i.e. where each has a different data distribution) such that the features the model learns are generic and robust across all said datasets. This is particularly important when there is no large-scale dataset for learning a specific task, but instead where several smaller datasets exists that have been created in different settings. Here, the authors develop a multi-domain learning pipeline for the task of identifying people who move between different CCTV camera feeds. Domain biases will result in certain neurons in the CNN being effective for one domain but not another. As such, the authors present a new dropout framework — Domain Guided Dropout — that assigns each neuron a specific dropout rate for each domain according to its effectiveness on that domain. Improvements on experimental datasets of person re-identification reached up to 46% above state of the art.
ICML2016: Benchmarking Deep Reinforcement Learning for Continuous Control, Cal Berkeley and Ghent University. Reinforcement learning, where an agent learns actions it should take in a given environment to maximise a cumulative reward, has seen loads of progress by leveraging deep learning for feature representations (vs. hand-engineering). Here, the authors present a new standardised and challenging testbed for evaluating algorithms in the continuous control domain, where data is high dimensional and model-free methods are often used. The framework consists of 31 continuous control tasks, ranging from basic to locomotion to hierarchical and will hopefully help researchers understand the strengths and limitations of their algorithms. Side note: here’s a great talk by Pieter Abbeel, author of the paper, presenting at Re.Work’s Deep Learning Summit in SF last quarter.
ICCC2016: Generative Choreography using Deep Learning, Peltarion and The Lulu Art Group. Here, the authors break ground on using end-to-end generative deep learning models for creating choreography. The authors use five hours of contemporary dance motion captured using the Microsoft Kinect v2 sensor tracking 25 joints to produce 13.5 million spatiotemporal joint positions in 3D. Using this data for training, the authors show that their network can output novel choreographies that demonstrate a progressive learning of increasingly complex movements (videos!). Creative AI making progress.
OpenAI have release their first public contribution in the form of the OpenAI Gym, which is a suite of environments for developing, comparing and reproducing results in reinforcement learning research. It includes algorithmic, Atari and classic control environments amongst others. Interesting that it lined up quite nicely with the Berkeley work above!
Speaking of simulation environments, have you ever wanted to play around with the architecture and parameters of a neural network? You’re in luck! Knock yourself out with this lovely browser demo!
Venture capital financings and exits
A slower period for investments ($36m announced in 26 deals) and 2 exits:
Eversight, which helps consumer goods brands and retailers to apply digital A/B testing to traditional retail using predictive analytics and behavioural economics, raised a $14.5m Series B led by Sutter Hill Ventures and Emergence Capital Partners.
DigitalGenius, a London-based developer of deep learning conversational agents for customer service, raised a further $4.1m in seed capital from investors including Bloomberg Beta and Salesforce Ventures. On the topic of bots, Betaworks in NYC have launched Botcamp, a program to accelerate the bot ecosystem (thanks Andrei @ Accel for pointing out!). *Are you bullish, bearish or neutral on bots becoming large independent companies?
Crosswise, a company based in Tel Aviv that identified users across multiple devices by building a probabilistic device map for marketers, was sold to Oracle for $50m. The business was founded in 2013, employed 19 FTE and raised $5.6m from Israeli investors Horizons Ventures, Giza VC and Pitango VC.
---
*Is Google synonymous with search? In spoken conversations, the term ‘Google’ features 52x per million words. That’s more frequent than the words ‘clever’, ‘fridge’, ‘eggs’ or death ‘death’. In fact, the phrase ‘search the Internet’ isn’t even present amongst 5m words from a conversational database!
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.