Your browser is not supported. Please update it.

20 April 2018

SenseTime hits historic valuation, as China aims for AI dominance

China’s relentless campaign to be the dominant country in AI continues to escalate and create the world’s most valuable startups in the field with several worth over $1 billion now. Top of the list comes SenseTime, which specializes in facial recognition technology and has just raised $600 million in a new funding round led by Alibaba, China’s answer to Amazon. This takes SenseTime’s value up to $4.5 billion by some counts, while rivals are also ratcheting up, with Shanghai based Yitu Technology now ranked as the second biggest AI start-up at $2.4 billion.

While SenseTime is China’s biggest start-up dedicated to AI, Didi Chuxing, the ride hailing company that bought out Uber in the country with 450 million users across 400 cities in the country, is over 12 times more valuable still at $56 billion. It has just raised $4 billion in additional funds much of which it says will be invested in AI research associated with autonomous driving and safety.

There are two reasons for betting on China to succeed in AI, despite a history of lagging well behind the US especially in both algorithmic innovation and cutting-edge hardware design, both critical for the field. One is the huge commitment made by the government, as well as by its big tech firms, notably Tencent Holdings, Baidu and Alibaba. The government recently announced it will build a $2.1 billion AI technology park in Beijing’s western suburbs in the latest in a spate of such investments. There are signs it is outspending the US which is investing around $1.2 billion a year on unclassified military AI research.

The second factor creating momentum is perhaps even more significant, the sheer amount of valuable data available for training AI and Machine Learning systems as well as for applications, resulting both from the country’s scale with 1.4 billion population and lack of scruples over privacy that impede development in the west. Data is the lifeblood of AI and is already giving China a leg up, especially over facial recognition and surveillance. The field is less about fundamental technology at the mathematical level and more about optimization, tuning and engineering – as one wag put it at an AI conference recently the last really serious innovation in the field was made in the 18th century by an English monk called Bayes who invented the rules of conditional probability and inference. While that is somewhat disingenuous it is true the Chinese are at the forefront of performance optimization and tuning to specific problems, notably in convolutional neural networks applicable to image and pattern recognition.

The development of convolution networks was instrumental in recent progress in image recognition. Neural networks comprise nodes or “neurons” connected in layers by links that have associated weights that allow training for specific tasks to take place. Each neuron is fully connected to neurons in the previous layer from which it receives input and in turn is connected to neurons in the next layer to which it delivers outputs, all weighted. The problem with image recognition is that there are a large number of links between successive layers each with potentially different weights, so execution quickly becomes computationally intensive at large sizes and high resolutions.

This means that such neural networks do not scale well given that an image of say 500 x 500 pixels would generate 250,000 inputs and potentially over 100 million parameters. They also suffer from overfitting whereby the model attempts to over-analyze and take too much account of sparse outlying data points which can generate false negatives. Two images of a given face under say different lighting conditions might not be recognized as the same person.

Convolutional neural networks take note of the fact it is dealing with images and constrain the search problem accordingly so that neurons of one layer are only connected to a small corresponding region of the layer above and below. They exploit the fact that two pixels are more likely to be similar if they are close together in the image than if they are far apart. This allows the system to delete large numbers of connections between neurons corresponding to distant pixels since these links are less likely to be relevant. In this way computational complexity is reduced by orders of magnitude for large images, at the expense of loss of generality so that the models are confined to image recognition. This is one reason why dedicated AI chips are being developed welding these algorithms to hardware for even greater performance improvements.

Chinese companies have been very active here at the R&D level, with Megvii, a 7-year old Beijing based start-up now worth about $1 billion, having published recent papers on methods for improving efficiency of what it calls deep convolutional neural networks (DCNN) even further by slimming down the input variables in size but not number. DCNNs simply have even more layers to give greater sensitivity in training but at the cost of generating still more parameters and threatening to overwhelm either network bandwidth or the resources of embedded processors in local processing.

As Megvii noted, a lot of R&D has already been performed to reduce computational effort on both general-purpose and specialized hardware with some success. Now Megvii has published results suggesting that the storage and computational burden associated with executing DCNNs can be cut significantly depending on the scale of the task by reducing the number of bits used to represent each number in the calculations. These are often small numbers that can be represented in just one or two bits. As the paper points out the complexity of say a multiplier in the execution is proportional to the square of the bit widths, so there is potential for great increases in the speed at which inferences can be drawn, as well as enabling models to be implemented in resource constrained processors suitable say for smartphones.

The approach is not confined to say facial recognition but could be applied to other areas of Machine Learning such as identification of objects which need not be images and also segmentation of instances. The latter, sometimes known as semantic segmentation, is itself crucial to deeper image processing by enabling identification of individual objects within a frame, which could be items of food on a plate or the contents of a still life painting. Development on this front is crucial for many fields, including autonomous driving in distinguishing between objects that must be avoided at all costs and those that can more safely be run over.

This example again highlights how AI related research is not so much about fundamental mathematics or statistics but practical ideas over how to implement specific tasks efficiency.

It is also worth noting that this idea of AI being a battleground between China and the west is misleading because there is a huge amount of collaboration taking place. Scruples over China’s disregard for privacy issues or even concerns over possible synergy with military applications tend to be overlooked in the heat of competition between enterprises as opposed to countries. So we find that Volkswagen has invested $250 million in setting up a joint venture with Beijing-based Mobvoi to develop voice recognition and language processing technologies for its vehicles. Mobvoi, now worth about $1 billion, is also backed by Google, having been founded by some of the latter’s’ ex staff. So even though Google is blocked in China this has not deterred it from investing there.

There is also the Hong Kong connection which is forging links between China and western companies in AI. Just over the water from Hong Kong the Shenzhen area has been dubbed China’s silicon corridor and is attracting high tech investment at least on a par with the Shanghai and Beijing areas. China allows Hong Kong to operate at arm’s length almost as a separate state under its “one country two systems” strategy and this encourages inward investment and collaboration with foreign companies that feel more comfortable operating there.

The South Korean internet portal operator Naver has just opened a research center specialized in artificial intelligence, in cooperation with Hong Kong University of Science and Technology (HKUST). This is focusing on the growing demand for AI services hosted in the cloud, which at this stage tend to be confined to prototyping and evaluation of different machine learning models, rather than full blown operational applications. It is possible AI services could extend to training of systems against full data sets but bandwidth might be an issue for some applications given the amount of data generated. Naver plans to extend the scope of AI services to areas such as intelligent mapping and location-based applications, as well as handwriting recognition where it is working in China with Mandarin scripts.