Every day, Amazon and its customers use artificial intelligence to make all kinds of processes more efficient, from delivering products to customers to generating real-time NFL game statistics. Why not apply the same technology to accelerate progress toward the United Nations’ sustainable development goals for 2030?
That’s the idea behind the Oxford Machine Learning Summer School (OxML), now in its second year. During the two-week virtual event, which takes place Aug. 9 through Aug. 20, attendees will learn about machine learning techniques from more than 30 lecturers, including James Hensman, an Amazon principal scientist.
The role of lecturer is familiar to Hensman, who once taught machine learning and statistics at Lancaster University in the United Kingdom. Now based a few hours away in Cambridge, he works within Amazon’s Supply Chain Optimization Technology (SCOT) organization, where for the past year he has focused on algorithms and simulations to improve inventory management and capacity control. At OxML, he’ll deliver a half-day tutorial to students joining from around the world.
“It’s a real pleasure to be able to teach people new concepts and hopefully influence them to come and work on things that I think are really exciting,” Hensman said.
Using artificial intelligence (AI) to solve big problems
OxML is organized by AI for Global Goals, a mission-driven company that aims to broaden the AI talent base and draw connections between AI and efforts to fulfill the 17 United Nations sustainable development goals (SDGs) for the 2030 agenda, which include climate action, zero hunger, and affordable clean energy.
We believe that more democratic development in AI across different industries and geographies, as well as underrepresented segments of society, is important for fulfilling the pledge to leave no one behind…
“We believe that more democratic development in AI across different industries and geographies, as well as underrepresented segments of society, is important for fulfilling the pledge to leave no one behind, which is at the core of UN’s SDGs,” said Mona Alinejad, founder and CEO of AI for Global Goals.
Alinejad, a biomedical engineer, worked for more than a decade in the healthcare industry. At conferences, she noticed that the topic of AI was missing from discussions about the SDGs. She also saw that machine learning workshops tended to remain confined to the theoretical. She wants to see more of that theory being applied to real problems related to the SDGs.
“There are a number of areas in medicine and healthcare, from drug discovery to image processes to electronic health records, where AI can have a huge impact,” she said. She pointed to the fact that even in developed countries such as the U.S., thousands of deaths annually can be attributed to simple medical error, another issue machine learning can help address.
The first OxML school was held last summer with more than 350 participants from 70 countries. Of those, according to Alinejad, 40% were female and 60% were from underrepresented groups in AI. She wants to expand the roster of events and ultimately connect these emerging AI workers with organizations that are working on the SDGs.
In addition to SDG No. 3 — good health and well-being, which also was OxML’s 2020 focus — this year the school will cover additional topic areas such as AI for Good (climate action, sustainable cities, and ESG, or environmental, social and corporate governance). Participants will attend sessions devoted to specific topics, such as computer vision and natural language processing. Separate sessions will be devoted to how these techniques connect to problems within the SDGs.
This year, the OxML event received a few thousand applications from 118 countries for around 400 slots, Alinejad said. The organizers look for underrepresented groups across geography, gender, and industry, along with the right technical background — the target education level is postgraduate students. Registration fees are steeply discounted for full-time academics and students, as well as attendees from low- to middle-income countries.
“We have more than 5,000 AI talents in our network,” Alinejad said. “Our goal is to become a platform for education and development of tech talents for the Global Goals.”
The beauty of the Gaussian process
Hensman’s lecture will be on the Gaussian process, a machine learning method that can be used to quantify confidence in a prediction. He describes Gaussian processes as a way to treat functions where you have an input (say, an image) and an output (a label describing the image).
Kalman filters solve 1d Gaussian process models; but don’t work well in ml frameworks like #TensorFlow (too much overhead in the loop). We pushed the loop inside the op, enabling HMC and VI on generalised GP models. https://128.84.21.199/abs/1902.10078 pic.twitter.com/av5YWt5cIn
— James Hensman (@jameshensman) February 27, 2019
“Rather than just having one function that we twiddle, with a Gaussian process we’ve now got a plausible space of functions that could reasonably explain the data,” he said. “Then when we try to tag a new image, or when we try to predict what’s going to happen inside of a new simulation, you can say, this is how confident we are.”
In terms of global sustainability goals, and particularly the health-related one central to this year’s OxML, Hensman points to several scenarios where this type of prediction method could be useful. Intensive care units might be able to use it to forecast patient metrics such as oxygen rates. Doctors could use it to assess the potential outcomes of administering certain medicines. X-ray technicians might use it to attach a likelihood of a diagnosis related to images.
Hensman lectured on the Gaussian process at last year’s OxML, and Alinejad said it was one of the most engaging lectures within the program. While Hensman says the Gaussian process is often perceived as an ivory tower method with a lot of “nasty, messy mathematics,” his goal is to let students know it’s not nearly as hard as they think.
“It’s actually quite intuitive,” he said. “You should really start thinking, ‘Could I just be using a Gaussian process here, would it actually make my life simpler and better? Would I get better-calibrated responses?’”
From theory to practice
Hensman, who earned his PhD in mechanical engineering, joined Amazon in April 2020. He enjoys applying theories within his research papers to simulations that ultimately help optimize inventory placement.
For example, when looking to forecast inventory flows, it’s not useful to estimate a single number without accounting for uncertainty. A Gaussian-process-based method can help give a clearer picture based on the data.
“We don’t want to say, ‘It’s going to be X million cubic feet of stuff arriving in your fulfillment center tomorrow,’” Hensman explained. “We want to say, ‘The chance of exceeding X million cubic feet is 95%; the chance of exceeding Y million is 5%.’”
The Gaussian process method is just one of many machine learning techniques Hensman applies as part of his work on supply chain optimization. What appeals to him about working at Amazon, he said, is the opportunity to move beyond theory and “change the course of business through the work that I do.”
And at OxML this summer, he’ll be inviting more curious minds around the world to think about how they can apply machine learning concepts to change the course of sustainability initiatives around the globe.
