Didi, China’s Uber equivalent, has been testing out a new algorithm for assigning drivers to riders in select cities.
MIT Technology Review
The Algorithm
Artificial intelligence, demystified
Reinforcement ride-hail
12.11.2018
Hello Algorithm readers,

Didi is upgrading its ride-hailing service. Didi, China’s Uber equivalent, has been testing out a new algorithm for assigning drivers to riders in select cities. The dispatching system uses reinforcement learning (RL), a subset of machine learning that relies on penalties and rewards to get “agents” to achieve a clear objective. In this case, the agents are the drivers and the rewards are their payments for completing a ride.

The company’s current dispatching algorithm has two parts: a forecasting system that predicts how rider demand changes over time, and a matching system that assigns drivers to jobs based on those predictions. It has served the company well thus far, but it can be inefficient. If the patterns for driver supply and rider demand change, the forecasting model needs to be retrained to continue making accurate predictions.

Moving to an RL approach solves this problem by collapsing both parts into one: with every subsequent piece of data, the algorithm learns to dispatch drivers more efficiently. That allows it to keep evolving with changing supply and demand, without any need to retrain. A/B tests between the old and new algorithms in a handful of cities confirmed that the new one was indeed more efficient.

Didi now plans to roll this new dispatching system out on a city-by-city basis. Tony Qin, the AI research lead for the company’s US division, told me it will conduct A/B tests between its different algorithms for each location and use the one that produces the most efficient results. The RL algorithm may not always be the best one, Qin said. It largely depends on the city’s supply and demand patterns. The company is also developing another RL dispatching algorithm, with different agents and rewards, to add to its arsenal.


The chilling effect of tough borders on AI. Last week over 100 researchers with tickets to attend AI conference NeurIPS were absent due to visa denials or delays. Many of those researchers were coming to Montréal from African countries and were supposed to present their work at a Friday event called Black in AI. Their inability to do so undermines efforts to make the field and its products more inclusive, researchers present told WIRED.

This isn’t the first time that scientists have been stopped from participating in international conferences. In fact, it has been a longstanding problem for those coming from developing countries to the US and Canada. Now as the US tightens its borders, the issue is affecting many more. Last year MIT Technology Review struggled to get a major researcher to our own AI conference because of his Iranian heritage.

These growing barriers have be unsettling to a research community that thrives on international collaboration. The stakes aren’t just the inclusivity of AI technologies, which already have a serious bias problem, but also the overall pace of innovation, which will slow if the best researchers around the globe cannot share their ideas. Visa struggles at conferences could also be the canary in the coal mine for much tougher barriers to come. In fact, the US is already considering restricting the export of AI and making it harder for foreign talent to come work in domestic AI labs. Both could seriously dampen the momentum to advance fundamental breakthroughs.

In parallel, the AI research community has held steadfast to its commitment for collaboration. Chinese tech giant Baidu joined an AI consortium founded by US tech giants, for example, to partner on ensuring the technology’s safety and fairness. Another major AI conference decided to host their 2020 proceedings in Ethiopia to bypass repeat visa problems. The question is whether those actions will be enough to counteract the tightening restrictions on the exchange of ideas.

More from TR

Andrew Yang, a 2020 presidential candidate, on why he’s focused on automation and universal basic income: “I’m convinced it’s driving the social, economic, and political dysfunction we are seeing. The reason why Donald Trump is our president today is we automated away 4 million manufacturing jobs in Michigan, Ohio, Pennsylvania, Wisconsin, Missouri, and Iowa, all the swing states he needed to win and did win. And everyone who works in technology knows full well we are about to do the same to millions of retail workers, call center workers, fast food workers, truck drivers, and on and on throughout the economy.” Read more here.

Mic Check

This is our first issue of The Algorithm using a new email platform. If you’re experiencing any issues, please let us know at algorithm@technologyreview.com.

Are you a thought leader, innovator, and executive looking to make your opinion heard?

We're now accepting applications to the MIT Technology Review Global Panel. Learn how to share your expertise and opinions today.


Research

Hidden data. The biggest obstacle to applying machine learning in medicine is a lack of massive amounts of data. Both stiff regulations and patient reluctance make it difficult for AI researchers to gain access to medical data. This trade-off between privacy and medical advancement has become a hotly contested topic. To get around the issue, MIT researchers have developed a new technique called a split neural network: it allows one person to start training a deep learning model and another person to finish it.

The idea is hospitals and other medical institutions would be able to train their models partway with their patients’ data locally, then each send their half-trained models to a centralized location to complete the final training stages with their models together. The centralized location, whether that be Google or another company’s cloud services, would never see the raw patient data; they would only see the output of the half-baked model plus the model itself. But the hospitals would still benefit from a final model that is trained on a combination of every participating institution’s data.

Ramesh Raskar, an associate professor at the MIT Media Lab and paper co-author, likens this process to data encryption. “Only because of encryption do I feel comfortable sending my credit card data to another entity,” he says. Obfuscating medical data through the first few stages of a neural network protects the data in a similar way.

In testing this approach over other data-protected machine learning techniques, the research team found that split neural networks require significantly less computational resources to train and produce models with much higher accuracy.

BITS AND BYTEs

Smartphone location data is far from anonymous—and it’s being sold to advertisers
Location data from apps that provide location-based services can be easily de-anonymized to reveal intimate details. (NYT)

Alexa is getting better at understanding users through contextual clues
New self-learning techniques allow it to understand mistaken commands and implied meaning. (Verge)

AI is no longer a tool but the product itself
A new class of consumer apps are shifting to a fully AI-reliant approach. (Andreessen Horowitz)

Better data policies could make the US criminal justice system more fair
With the right accountability measures, automation could put an end to mass incarceration. (Harvard Kennedy School)

The Good Place reverses the worn trope of killer AI
Janet, an virtual assistant who tirelessly helps the people around her, depicts an optimistic narrative about how humans and AI can coexist. (WIRED)

Quotable

By 2030, most social situations will be facilitated by bots – intelligent-seeming programs that interact with us in human-like ways.

Judith Donath, faculty fellow at Harvard University’s Berkman Klein Center, on her predictions for how AI will change humans

Karen Hao
Hello! You made it to the bottom. Now that you're here, fancy sending us some feedback? You can also follow me for more AI content and whimsy at @_KarenHao.
Was this forwarded to you, and you’d like to see more?
New Call-to-action
You received this newsletter because you subscribed with the email address:
edit preferences   |   unsubscribe   |   follow us     
Facebook      Twitter      Instagram
MIT Technology Review
One Main Street
Cambridge, MA 02142
TR