News in artificial intelligence and machine learning

From 2nd June thru 11th July 2016. Referred by a friend? Sign up here. Want to share? Give it a tweet
July 11 - Issue #13

Nathan Benaich

An analytical digest of artificial intelligence and machine learning news from the technology industry, research lab and venture capital market.

From 2nd June thru 11th July 2016. Referred by a friend? Sign up here. Want to share? Give it a tweet :)
*We held our second annual Playfair AI Summit on research and applied AI last week, with 350+ attendees taking part. Talk slides are available here, photos here and videos are forthcoming!
*The next edition of the London.AI meetup, featuring Mapillary, Tractable and StatusToday is happening on July 21st. Register here.
*My top pick for this issue…It’s ML, not magic: simple questions you should ask to help reduce AI hype, by Stephen Merity of MetaMind-Salesforce.
This issue is a meatier 15 min read 🙈 as it covers 5 weeks of news 👍

Technology news, trends and opinions
🏆 Recognising achievements in AI
The Economist published a brilliant special report on Artificial Intelligence. Tom Standage opens by making the case that today’s world is fertile ground for AI to deliver on its promise. This time it’s for real.
WIRED open their Innovation Awards series by celebrating British leaders in AI, including the founders of Jukedeck, SwiftKey (acq. Microsoft), VocalIQ (acq. Apple), Blippar and Google DeepMind. I was fortunate to be involved in this project!
Four machine learning-enabled geospatial companies were elected as Technology Pioneers 2016 by the World Economic Forum: Mapillary (crowdsourced streetview), Orbital Insight (geospatial big data), Farmers Edge (precision agriculture) and PrecisionHawk (UAVs).
🎨 Creativity, design and theories of AI
Four announcements showcasing how creative applications of AI are picking up the pace:
  1. Following the launch of Google Brain’s creative AI project Magenta in May, the team released their first track. Compare it to this piece written by Jukedeck’s AI composer and you’ll have confidence that startups can indeed outperform the giants.
  2. A researcher at Goldsmiths in London trains a variational autoencoder deep learning model on all frames from the Blade Runner movie and then asks the network reconstruct the video in its original sequence as well as other videos the network wasn’t trained on.
  3. Researchers at NYU trained a recurrent neural network on scripts from movies including Ghostbusters, Interstellar and The Fifth Element, and asked the network to generate a novel screenplay. Watch the result acted out here!
  4. Work from MIT’s computer science and AI lab shows how to predict sound from a silent video. The algorithm takes video sequence as input and predicts a corresponding sequence of sound features. Next, it synthesises a waveform by matching these features with a database of impact sounds and transferring the best matches. Watching someone strike objects with a drumstick has never been so interesting :)
A designer’s guide to AI. Leveraging user centered design principles, the author rightly states that AI will enable designers to create bespoke experiences right out of the box for each user. Importantly, these experiences need to a) create emotionally-aware relationships with the user, b) respond to needs that haven’t yet been explicitly expressed, c) prevent negative emotional responses when a user is upset with an AI-caused result and d) be sensitive to sociology. A list of further reading resources is included. Thanks to Joe Thornton for sharing!
There’s been a resurgence of neuroscience-inspired AI architectures in the past few years, with Numenta being one of the leaders. Their VP of Research, Subutai Ahmad, argues that environmental sensory inference and behavior generation are intricately tied together, and critical for learning really general purpose representations. This contrasts with fitting models to data and outputting predictions.
An MIT student explains the ups/downs of AI through the lens of gradient descent, where a decreasing slope to the function that defines progress in AI implies that we’re reducing the difference difference between artificial and biological intelligence. He argues that deep learning can be a plausible substrate for future AI systems and that we should focus on solving tough problems to avoid another Winter.
💊 AI in healthcare
Google DeepMind announced their research project with Moorfields Eye Hospital in London aimed at the early detection of preventable eye disease (e.g. age-related macular degeneration and diabetic retinopathy). This work involves analysing optical coherence tomography scans of the retina.
Further details emerged on the DeepMind’s working relationship with the NHS, which outlined the scope for projects to improve clinical outcomes, patient safety and cost efficiency of the NHS. Head of Applied AI, Mustafa Suleyman, also penned his commitment to the NHS.
Microsoft suggests they’re able to detect (to some extent) early signs of pancreatic cancer in users of their Bing search engine. They claim to disambiguate queries that were caused by anxiety or linked to genuinely experiencing symptoms of the disease. Separately, the Chinese startup, iCarbonX, set up by the co-founder of the renowned Beijing Genomics Institute raised $200m to predict the onset of disease from genomic, medical and lifestyle data. This work highlights how we’re creating data assets that are descriptive of our health that didn’t exist in the pre-Internet era.
I’ve tracked close to a dozen new startups founded in the last 12 months applying deep learning to medical imaging, such as BayLabs, Imagia,, AvalonAI,, Kheiron Medical, Enlitic and others. This piece synthesises key technical challenges, publicly available datasets and approaches to problem solving.
🔏 Privacy and social implications AI
Apple plays feature catch-up with Google, namely on its photo tagging/search capabilities and predictive keyboard, but takes a view that privacy should come first. The Company is using on-device machine learning and differential privacy, a method to protect against deanonymizing efforts when querying statistical information about a sensitive dataset. This compares to Google’s stance of using cloud-based computation (deemed less safe) to power its machine learning features.
Microsoft CEO, Satya Nadella, outlines three key tenets of his vision for developing AI: 1) Augment human abilities and experiences instead of replace us, 2) Work to earn a user’s trust by solving privacy, transparency and security, 3) Technology should be inclusive and respectful of all users.
Following from this point, the New York Times features a piece on how algorithms perpetuate intrinsic biases in their training data, drawing on examples from the police force, image classification tasks and gender discrimination. While no solutions are presented, techniques which expose the reasons why algorithmic outputs were achieved in addition to careful creation of the training dataset would help greatly.
🆓 Open source, FB and GOOG 
Google release an open source implementation of what they call Wide and Deep Learning, a two-pronged model that seeks to emulate our ability as humans to memorise concepts and subsequently generalise this knowledge to new situations. This approach is suited for classification and regression tasks where input data points are sparse (e.g. food recommendation engine or search results). Paper here.
The Backchannel ran a great piece on the various initiatives at Google to infuse machine learning into all of the Company’s products and train up more of their staff. A new dedicated Machine Intelligence research group was also announced in their Zurich office.
Facebook’s Language Technology team, which forms part of Applied ML, was the subject of a recent expose by Forbes diving into its various initiatives. The team recently published their text understanding engine, DeepText, that is able to understand sentiment, intent and entities across more than 20 languages. They’ve also built a new multilingual composer to enable authors of posts on Facebook Pages to reach audiences speaker other languages using automatic machine translation.
On the topic of NLP, here’s an educational interview with the creator of OpenNLP on where the field is today, outstanding questions in computational linguistics and how machine learning and neural approaches fit in. As a fun aside, this project analysed sentences from Donald Trump and Hillary Clinton to learn characteristics to distinguish them. Trump is emotionally charged while Hillary is pragmatic and topical.
🚗 Driverless cars and other robots
Probably the cutest, most advanced and largely autonomous fleet of delivery robots in town, Starship Technologiesannounced their launch partners including Just Eat, Hermes and Pronto. I’ve seen one in action and can’t wait to cross more on the sidewalk soon!
Boston Dynamics published an incredible video of their newest creation, the SpotMini, which is all-electric and runs for 90 minutes. It can operate some tasks autonomously and is capable of climbing stairs, picking itself up and handling sensitive grasping tasks. No news on its pending sale since Toyota emerged as a likely buyer last month.
NVIDIA, arguably the leaders in GPU chips for deep learning, present their second generation water-cooled autonomous driving processing supercomputer (Drive PX 2). The system will process sensors, make control decisions and offer non-real time image analysis. NVIDIA will be a force to be reckoned with in an increasingly dense supplier landscape.
On the subject of driverless cars, the ethical framework for defining algorithms that make moral decisions is an outstanding question. Here, researchers show that study participants want to be passengers in vehicles that protect their riders at all cost while preferring that others purchase vehicles controlled by utilitarian ethics, i.e. sacrificing its passengers for the greater good.
Research, development and resources
The 33rd International Conference on Machine Learning, a major event on the AI calendar, took place last month. I encourage you to skim through two sets of summaries notes of the best talks and papers: this one and that one. David Silver of Google DeepMind/AlphaGo fame also ran a brilliant session on reinforcement learning, which you can learn more about in his recent blog post.
Here’s my selection of exciting applied AI research:
Concrete problems in AI safety, Google Brain, Stanford, UC Berkeley and OpenAI. Here, the authors define 5 practical research problems in AI safety (relating to reinforcement learning), defined as preventing unintended and harmful behavior that may emerge from poor design of real-world AI systems. These areas can be summarised as: 1) Safe exploration of an environment by an agent, 2) Robustness to changes in the distribution of input data, 3) The avoidance of unwanted impacts on the environment as a result of learning optimal policies, 4) Avoiding cheating behaviour to optimise a reward function, 5) Learning policies when feedback is expensive to give (and thus sparse).
Safely interruptible agents, Google DeepMind and The Future of Humanity Institute/Oxford. The authors focus on the question of defining the optimal interruptible policy for a reinforcement learning agent. Of note, they show that their Q-learning algorithm, which played Atari games, is safely interruptible. Moreover, they make the case for ensuring that interruptions do not jeopardise the agent’s goal of behaving optimally in its environment. The agent should therefore behave optimally under the assumption that there won’t be future interruptions to its exploration of the environment.
OpenAI publish new work (papers + explanation) extending generative adversarial models (GAN), a neural network architecture proposed by Ian Goodfellow in 2014 that uses a generator to produce data (e.g. an artificial image) from a random input and a discriminator receiving an input from either the generator (the artificial image) or a real dataset (a real image) and is tasked with calling out artificial data. The discriminator must maximise the probability of assigning the correct label to the input (1=real image, 0=artificial image), while the generator optimises to fool the discriminator. The other important feature of the adversarial training method is that the network learns its own cost function instead of requiring it to be crafted by the engineer, meaning that the network learns its own rules for which outputs are good or bad. As such, we’re able to create a model capable of generating data that approaches reality (i.e. images that look real but are actually computer-generated) by learning hierarchical feature representations without supervision. The OpenAI work introduces techniques for making GAN training more stable. They also present a semi-supervised learning approach where the discriminator outputs an image label such that the network can outperform image classification tasks with only 10 labeled examples per class (99.41% accuracy on MNIST handwritten digit dataset).
Synthesizing the preferred inputs for neurons in neural networks via deep generator networksUniversity of Wyoming, Geometric Intelligence and University of Freiburg. With few exceptions, neural network models are currently black boxes that obfuscate the reasoning underlying their outputs. However, being able to explain why a network classified an image in a certain way or elected not to approve a financial decision is key to their adoption in fault-intolerant domains. The key to shedding light on the inner workings of deep neural networks is to discover the input data that highly activates a specific neuron in the network, a process called activation maximisation. Here, the authors use a deep generator network to perform activation maximisation whereby a generative model outputs a synthetic image that looks as close to images from the ImageNet dataset. Thus, one can understand why a neural network classifies a dog as a dog because the input synthetic input image that highly activates the neurons responsible for the classification is that of a dog.
Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural networkMagic Pony Technology and Imperial College London. Curious about the technology Twitter paid $150m for? In this work, the London-startup address the problem of upscaling the quality of a single image or video from low resolution (LR, which is blurry, downsampled and noisy) to high resolution (HR, which is sharp). This is topical for high definition television streaming, medical and satellite imaging, which are particularly bandwidth and computationally expensive. Existing methods for super-resolution (SR) rely on having multiple LR images of the same scene or instead use computationally expensive convolutional neural networks (CNNs)-based methods that increase resolution before enhancements. Here, the authors show that by upscaling from LR to HR at the end of the network, feature extraction occurs in LR space from which HR data is super-resolved. This gives rise to 10x speed and performance compared to the state of the art CNN approaches, making it possible to run super resolution HD videos in real time on a single GPU. Awesome!
Venture capital financings and exits
57 financing events (55 VC and 2 PE) totalling $500m, including:
  • Zoox, the stealth startup developing both hardware and software for an autonomous taxi fleet, raised a $200m Series A at a $1bn valuation from Aid Partners in Hong Kong, Lux Capital and DFJ.
  • Orbital Insight, which creates computer vision based geospatial-analysis software, raised a $20m Series B led by Google Ventures and including Bloomberg Beta, Lux Capital and Sequoia. Very neat work here - check out this recent talk by founder here.
  • Darktrace, the ever growing London-based cybersecurity company using machine learning to detect novel threats, raised $65m led by KKR with participation from Softbank and Summit Partners. This values the 3 year old company at $400m post-money.
7 exit events worth $183m in total announced consideration, including:
  • Magic Pony Technology, the London-based startup developing machine learning technologies for visual processing, was acquired by Twitter for $150m. The startup raised $6.3m in two rounds from Octopus Ventures, Balderton Capital, Entrepreneur First and angels. It employed a dozen staff and filed 20 patents in its two year life, marking a significant exit for the European AI ecosystem. Congrats all!
  • Moodstocks, a French computer vision company selling a mobile SDK for image recognition, was purchased by Google for an undisclosed sum. The company was half the size of Magic Pony and founded in 2008, having raised less than $1m in its life. 
Anything else catch your eye? Do you have feedback on the content/structure of this newsletter? Just hit reply! 
I’m actively looking for entrepreneurs building companies that build/use AI to rethink the way we live and work.

Do you enjoy receiving this? Please share it:
Carefully curated by Nathan Benaich with Revue
If you were forwarded this newsletter and you like it, you can subscribe here.
If you don't want these updates anymore, please unsubscribe here.