Page 78 - The Economist

Science and technology
78 The Economist December 9th 2017

Also in this section
79 Cleaning rivers with mussels
79 Gender in academia
80 Portable DNA sequencing

For daily analysis and debate on science and
technology, visit
Economist.com/science

Artificial intelligence machine-vision algorithms to the video
Algorithm is gonna get you 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 goodsto the bestadvantage. With the
rightalgorithms, Targetcould automate the
process and run it in real time.
LONG BEACH, CALIFORNIA Many firms were also putting on a
Machine learning’s big shindig
show as part of the battle for AI talent.
ORPORATE conferences still suck.” accelerate the pace of research into a form They included Mercedes-Benz, a first-time
“CSo read the T-shirt sported by Ben ofmachine learning known as deep learn- sponsor, which is trying to recruit data sci-
Recht, a professor at the University of Cali- ing, leading to huge advances in image rec- entists to work on its autonomous cars.
fornia, Berkeley, ashe collected an award at ognition in 2012. Deep learning, which The German producer is already some
the Neural Information Processing Sys- stacksmanyneural networkson top of one way down the road, with Rigel Smiroldo,
tems(NIPS) conference thisweek. DrRecht, another to learn the features of giant data- the firm’s machine-learning boss in North
pictured above in lecture mode, was prot- bases, now powers the image-processing America, happy to recite how the E-class
esting against the flood of corporate mon- operationsoffirmslike Facebookand Goo- Mercedes he drove to NIPS handled 250
ey pouring into NIPS, aping the words Kurt gle. As machines, trained with heaps of miles of highway driving without him
Cobain wrote on a T-shirt when he ap- data to develop cleveralgorithms, have be- needingto intervene.
peared on the coverofRollingStone in 1992. come capable of carrying out more and
“It’s not an academic conference any- more tasks, so interest has grown. Google Yes, no and now, maybe
more,” Dr Recht says wistfully, perched in wassponsoringNIPSby2010, and thisyear Mr Smiroldo does put his finger on one of
the Californian sun on the steps of the all ofthe world’slargesttech firmscould be the main trends at this year’s NIPS: the
Long Beach Convention Centre. He com- found on the sponsorsheet. merging of Bayesian statistics with deep
plains that folk would rather go to cor- For the 7,850 attendees, the big draw is learning. Instead of algorithms presenting
porate-sponsored parties these days (In- the algorithms presented in halls heaving deterministic “yes” or “no” results to que-
tel’s featured Flo Rida, a rapper), than with mostly male bodies (90% of the au- ries, new systems are able to offer up more
poster sessions. AI, it seems, is the new thors ofNIPS papers were male this year, a probabilistic inferences about the world.
rockand roll. gender imbalance widely found in sci- This is particularly useful for Mercedes-
NIPS began in 1987 as a humble little ence—see later story). They hang on every Benz, which needs driverless cars that can
conference on an obscure branch ofmach- word of AI wisdom imparted by luminar- handle tricky situations. Instead of an al-
ine learning called neural networks. It ies from Google and Microsoft; pore over a gorithm simply determining ifan object in
spent the first 13 years of its life in Denver, dizzying number of advances (laid out in the road is a pedestrian or a plastic bag, a
then moved to Vancouver for a decade. It more than 670 published papers) from the system using Bayesian learning offers a
used to be a quiet affair, with a few hun- likes of Facebook, DeepMind (a unit of more nuanced view that will allow AI sys-
dred mathy computerscientists coming to- Google) and Tencent; and devourstories of tems to handle uncertainty better.
gether to explain how they had solved novel ways to train machines to perform Netflix already uses data science to rec-
some abstract problem in a new way. useful tasks. ommend shows to its subscribers. Nirmal
Then, at the 2003 conference, Geoffrey Those storiescome notjustfrom the big Govind, who develops algorithms at the
Hinton, a British polymath, and a cabal of names of technology, but also from more firm, was on the lookout at NIPS for new,
AI researchers founded the Neural Com- old-fangled companies, such as Target, a improved versions that can handle imag-
putation & Adaptive Perception (NCAP) bricks-and-mortarAmerican retailer. Brian ery and video. The firm is particularly in-
working group. As a proponent of neural Copeland, one of the firm’s data scientists terested in automating the generation of
networks, DrHinton and the group helped in Minneapolis, says he is trying to apply promotional material around its original 1