A steady flow of stories in the Chinese press has signalled that the era of artificial intelligence is advancing quickly in the world’s most populous country. The latest examples include facial recognition technology used in the arrest of suspected criminals at three different concerts of the Hong Kong singer Jacky Cheung (see WiC410), and the revelations from a government scientist that one in every four schools is testing a machine that uses artificial intelligence to mark pupils’ work.
For insights from one of the leading companies in the sector, WiC met Matt Scott, the chief technology officer of Malong Technologies, on the sidelines of HSBC’s China Conference last month.
Shenzhen-based Malong, which was founded in 2014 and has since been backed by Softbank, is a world leader in computer vision and image recognition. WiC asked Scott for more on how artificial intelligence is being de-ployed, and whether China’s AI pioneers have already reached unicorn status.
Where are artificial intelligence techniques being used most?
Task-based AI, otherwise referred to as Artificial Narrow Intelligence (ANI), is already a feature in a number of industry verticals and in many cases it has reached human-level performance or exceeded it. ANI is a practical form of AI today that aims to enable machines with the capability to perform specific cognitive tasks incredibly well, as opposed to Artificial General Intelligence (AGI), which seeks to provide machines with the capability to experience consciousness. ANI and AGI are often conflated with one another.
Malong started out in e-commerce, where we developed AI for the task of product recognition. This powers many retail websites and apps in China and around the world. That is, the task is to enable machines to be able to “see” products like people can, for visual search scenarios and extracting attributes for classification scenarios. For example, China is by far the biggest exporter of textiles, but it can be difficult for textile firms to convey the differences in patterns or styles across their product ranges. That makes visual search very important for their customers. Malong’s computer vision API called ProductAI has over 50% market share across platforms where customers reference pictures of the kind of textiles they want and look for matches in the company catalogues. We are applying the same concept across a wider spread of goods in the worlds of fashion, furniture, wine and all forms of consumer-packaged goods.
We’ve since taken our product recognition technology to the next level by enabling it to operate not just in the cloud for e-commerce, but on local hardware to power “RetailAI” scenarios in physical stores as well. For example, to enable cashier-less shopping scenarios, automatic inventory checking, and visual audit inspection.
Taking visual product recognition even further, we’ve applied it to other interesting offline scenarios, where the focus is not necessarily about reducing labour costs, but on improving quality and the user experience. One such example is product recognition in manufacturing, where automatic quality inspection is applied to textiles. Previously, only random fabric samples would be checked, now with AI, we can check vastly more samples in a fraction of the time. Another example is baggage screening. Often, there are lines at screening points because security personnel take a long time to scrutinise the baggage. AI-enabled machines can work much faster, bringing down queuing time. Additionally, they don’t get sleepy or distracted.
How about ‘new retail’, which has been getting a lot of media coverage?
We have trials underway with a number of retailers in China. One area is ‘smart cabinets’, which offer a kind of combination of mini-stores and vending machines. The cabinets are unmanned and can be installed on inexpensive real estate in companies and communities. As a customer, you scan a QR code from WeChat, and the cabinet opens up. You pull out the product you want, and then you get charged online. We have cameras inside the cabinets, powering the product recognition algorithms that identify the goods that are being purchased.
Healthcare is another important area for Malong. There are lots of ways that artificial intelligence can help doctors be more productive or speed up their decision-making. One example is cases of ischemic strokes, where it is critically important to identify blockages of blood flows in the brain quickly. Doctors need time to review the CT scans and figure out what is going on. Our technology can analyse the images instantly and flag the problem to the medical professionals.
Scoliosis, or curvature of the spine, is another condition where we have been contributing, although in this case in recommending treatment after diagnosis. Our technology looks at images of the spine and calculates what the best course of treatment should be. Of course, doctors can do this themselves but automating the process allows treatment decisions to be taken more efficiently. We are currently in clinical studies in Shenzhen.
What makes Malong stand out from the crowd?
Most artificial intelligence firms are utilising a technique called ‘supervised learning’. Essentially, it relies on highly curated datasets in which companies are shifting the complexity of their business model away from the programming into data management.
What many people don’t realise is that many of these firms have armies of people behind the scenes, making the datasets more intelligible before they feed them to the algorithms. The idea is to make sure that the data is labelled properly so that it can be processed effectively.
Malong has developed a breakthrough in a more sophisticated approach called ‘weakly supervised learning’. Our goal is to take the data largely as it is and generate high-performance outcomes without the need to curate it to the same level.
Last year we won first prize in a Google-backed worldwide computer vision competition called WebVision, against more than a hundred other institutions and companies, including billion-dollar firms. Our model was more accurate by a wide margin, achieving human-level performance from noisy data.
How do you do it?
Because of our team, our culture, our experience and our assets. With regards to the team, we have an in-house research group of world-class scientists in artificial intelligence and computer vision hailing from top institutes such as Microsoft Research and the University of Oxford. We’ve taken about four years of research and development to create robust, high-performance algorithms that can learn general representations of objects, such as prod-ucts in retail, or lesions in medical image analysis, from noisy training data. We have discovered new deep learning training algorithms inspired by the way human beings learn, from children to adults.
Malong’s platform allows for more accurate matching and it deals more effectively with chaotic data. It also outperforms in situations where there is a regular flow of fresh or new data, which is important because retailers don’t like situations in which a new model has to be trained each time a new product is released.
We plan to publish more about our techniques at the leading computer vision scientific conferences.
Bigger picture, there’s a lot of talk about rivalry between China and the US for leadership in the world of AI? Is it accurate?
I can’t really talk about geopolitics but I can tell you about the research sphere, where the scientists and professors are doing the work in moving the industry forward.
In China some great research is coming out of Tsinghua, for example, where Malong shares a lab with the university. In the United States, places like Berkeley are the leading players and Oxford in the UK is also very respected. Malong is fortunate to have talent and collaboration with all of these institutions.
In this community the mood is collaborative. The scientists attend the same conferences and they publish their research. Different nationalities are working together and achieving great things. Amongst this audience, there isn’t much talk of the so-called ‘arms race’ in artificial intelligence.
In a similar way, our clients just want to know how our platform will benefit their customers. And in sectors like healthcare, the benefits will extend across borders, including the imaging projects we are running with hospitals in Shenzhen. The work there is going to develop models that help millions of patients around the world, not just in China.
Are more AI firms going to achieve ‘unicorn’ status?
The great thing about the industry is that there are lots of opportunities for smaller companies. The pie is so big because there are so many ways that artificial intelligence can be applied. A number of companies have hopes of becoming unicorns.
Another feature in the industry is that its leading firms can be built around small clusters of people. In the world of software engineering, companies had to be huge to be successful, with thousands and thousands of staff. But AI is based on scientific research, where scale, or access to finance, doesn’t have quite the same significance. You don’t need to be Facebook or Google to survive or thrive. You can anchor a pioneering company around a small group of experts.
Nor is it just about the money. If that were the case, the banks would be the best at AI, although there’s still a baseline for attracting the right people and you have to pay very high salaries for top talent.
What’s interesting is that it isn’t a case of Chinese firms paying less than their American counterparts – in some cases people in China are being paid more.
However, the main point is that smaller firms can compete if they are smart enough, especially in situations in which the leading talents often want to work with other outstanding people. At Malong we go head-to-head with some of the world’s biggest firms and we still win the business.
What’s Malong’s exit strategy? Aren’t the best firms being acquired by the BAT?
It has been more common for companies like Alibaba and Tencent to make strategic investments in start-ups, rather than buying them outright.
In my view, the AI landscape is more diverse in China than in the United States, with a larger number of independent operators. In the American market there is more consolidation and the buyouts happen earlier in the cycle, often at $50-100 million valuations. Every year I turn up at industry conferences in Silicon Valley and I notice that some of our competitors have disappeared. Usually they’ve been bought by the bigger players and absorbed into their businesses.
In China the AI firms see their financial destinies differently. Most of the start-ups that emerged at a similar time to Malong are still around and we share their ambitions of building bigger companies with bigger valuations.
Speaking for Malong, we have a business goal and a technology goal. In technology terms we want to be the leading brand for “RetailAI” and “MedicalAI” worldwide. And commercially, our objective is to get to IPO. Fortunately, changes in the way that technology companies are starting to get access to the financial markets are making that more of a possibility.
© ChinTell Ltd. All rights reserved.
Sponsored by HSBC.
The Week in China website and the weekly magazine publications are owned and maintained by ChinTell Limited, Hong Kong. Neither HSBC nor any member of the HSBC group of companies ("HSBC") endorses the contents and/or is involved in selecting, creating or editing the contents of the Week in China website or the Week in China magazine. The views expressed in these publications are solely the views of ChinTell Limited and do not necessarily reflect the views or investment ideas of HSBC. No responsibility will therefore be assumed by HSBC for the contents of these publications or for the errors or omissions therein.