Science & technology | Algorithm is gonna get you

All the buzz at AI’s big shindig

Machine learning’s big event

|LONG BEACH, CALIFORNIA

“CORPORATE conferences still suck.” So read the T-shirt sported by Ben Recht, a professor at the University of California, Berkeley, as he collected an award at the Neural Information Processing Systems (NIPS) conference this week. Dr Recht, pictured above in lecture mode, was protesting against the flood of corporate money pouring into NIPS, aping the words Kurt Cobain wrote on a T-shirt when he appeared on the cover of Rolling Stone in 1992.

“It’s not an academic conference anymore,” Dr Recht says wistfully, perched in the Californian sun on the steps of the Long Beach Convention Centre. He complains that folk would rather go to corporate-sponsored parties these days (Intel’s featured Flo Rida, a rapper), than poster sessions. AI, it seems, is the new rock and roll.

NIPS began in 1987 as a humble little conference on an obscure branch of machine learning called neural networks. It spent the first 13 years of its life in Denver, then moved to Vancouver for a decade. It used to be a quiet affair, with a few hundred mathy computer scientists coming together to explain how they had solved some abstract problem in a new way.

Then, at the 2003 conference, Geoffrey Hinton, a British polymath, and a cabal of AI researchers founded the Neural Computation & Adaptive Perception (NCAP) working group. As a proponent of neural networks, Dr Hinton and the group helped accelerate the pace of research into a form of machine learning known as deep learning, leading to huge advances in image recognition in 2012. Deep learning, which stacks many neural networks on top of one another to learn the features of giant databases, now powers the image-processing operations of firms like Facebook and Google. As machines, trained with heaps of data to develop clever algorithms, have become capable of carrying out more and more tasks, so interest has grown. Google was sponsoring NIPS by 2010, and this year all of the world’s largest tech firms could be found on the sponsor sheet.

For the 7,850 attendees, the big draw is the algorithms presented in halls heaving with mostly male bodies (90% of the authors of NIPS papers were male this year, a gender imbalance widely found in science - see article). They hang on every word of AI wisdom imparted by luminaries from Google and Microsoft; pore over a dizzying number of advances (laid out in more than 670 published papers) from the likes of Facebook, DeepMind (a unit of Google) and Tencent; and devour stories of novel ways to train machines to perform useful tasks.

Those stories come not just from the big names of technology, but also from more old-fangled companies, such as Target, a bricks-and-mortar American retailer. Brian Copeland, one of the firm’s data scientists in Minneapolis, says he is trying to apply machine-vision algorithms to the video feeds from the cameras in Target’s stores. Retailers employ behavioural experts to watch such videos so they can work out how people use their stores and where to place goods to the best advantage. With the right algorithms, Target could automate the process and run it in real time.

Many firms were also putting on a show as part of the battle for AI talent. They included Mercedes-Benz, a first-time sponsor, which is trying to recruit data scientists to work on its autonomous cars. The German producer is already some way down the road, with Rigel Smiroldo, the firm’s machine-learning boss in North America, happy to recite how the E-class Mercedes he drove to NIPS handled 250 miles of highway driving without him needing to intervene.

Yes, no and now, maybe

Mr Smiroldo does put his finger on one of the main trends at this year’s NIPS: the merging of Bayesian statistics with deep learning. Instead of algorithms presenting deterministic “yes” or “no” results to queries, new systems are able to offer up more probabilistic inferences about the world. This is particularly useful for Mercedes-Benz, which needs driverless cars that can handle tricky situations. Instead of an algorithm simply determining if an object in the road is a pedestrian or a plastic bag, a system using Bayesian learning offers a more nuanced view that will allow AI systems to handle uncertainty better.

Netflix already uses data science to recommend shows to its subscribers. Nirmal Govind, who develops algorithms at the firm, was on the lookout at NIPS for new, improved versions that can handle imagery and video. The firm is particularly interested in automating the generation of promotional material around its original shows and finding ways to make that material more engaging.

Besides fundamental algorithms which firms hope to apply to their own operations, NIPS is also home to applied research, particularly in health care and biology. Becks Simpson from Maxwell MRI, a startup from Brisbane in Australia, showed a way to combine magnetic resonance imaging with deep learning to improve the diagnosis of prostate cancer. Elisabeth Rumetshofer from Johannes Kepler University Linz presented a system that could automatically recognise and track proteins in cells, helping to illuminate the underlying biology. A team from Duke University in North Carolina had used machine learning to detect cervical cancer automatically using a pocket colposcope, to the same level of accuracy as a human expert. Some used AI to mine doctors’ notes to estimate the chances that a patient will be readmitted to hospital, to categorise and understand the allergic reactions of children and to model the geographic distribution of naloxone, which can help block the effects of opioids, in order to get a better grip on the use of such drugs.

Other applications ranged from researchers at the Federal University Lokoja in Nigeria trying to use machine learning to identify potential child suicide bombers to the Donders Institute in the Netherlands presenting a system that can reconstruct pictures of faces that a person sees simply by scanning their brains. Google researchers used machine learning to hide a complete image inside another picture of the same size. What they might do with that remains to be seen.

New hardware for machine learning was on display, too. At its party Intel unveiled its latest chip dedicated to solving AI problems. NVIDIA, a chipmaking rival whose share price has increased ninefold in the past three years thanks to sales of its graphical-processing units for deep learning, displayed its latest wares. Graphcore, a British startup, caused particular waves. It presented benchmarks for its chip’s performance on common machine-learning tasks that tripled speeds for image recognition and delivered a claimed 200 times improvement over NVIDIA for the kinds of machine learning required for speech-recognition and translation applications.

Among older hands at NIPS, especially those who can remember its origins, there is a sense that the corporate obsession with machine learning will not last. They should not be so sure. The systems being developed are just beginning to be a broadly useful technology, and new algorithms presented at the conference are likely to be adopted rapidly. Powerful computers and large volumes of data lie waiting for exploitation. The world’s most valuable companies have grasped the power of machine learning, and they are unlikely to let go.

This article appeared in the Science & technology section of the print edition under the headline "Algorithm is gonna get you"

The choice that could save South Africa, or wreck it

From the December 7th 2017 edition

Discover stories from this section and more in the list of contents

Explore the edition

Discover more

Antarctica, Earth’s largest refrigerator, is defrosting

The world must pay more attention to its southern pole

Killer whales deploy brutal, co-ordinated attacks when hunting

Their techniques are passed down through the generations


A new generation of music-making algorithms is here

Their most useful application may lie in helping human composers