There is no consistent definition of digitization. Every industry and every department perceive it and react differently. How would you define digitization with respect to your industry?
Michael Wei: Digitization has been the major shifting force for the last 4 decades, and it will continue to be a critical force in the next several decades ahead of us. The scope and impact are substantially wider than what we perceive, that’s where I think the inconsistency comes from – each player stems from its own roots to look at digitization and has reached a definition from a partial view. Actually we are all part of a giant transformation, “digital world is eating up physical world” – enterprise, individual, activity, relationship, emotion, etc., almost every entity around us is getting through some degree of digitization. Personally, I am particularly interested in the digital innovation in areas of user experience and productivity improvement. Taking a specific example, I believe the future digital world will provide us a completely new manner to experience with, starting from innovative Human Machine Interactive (HMI). I personally have a strong interest in the natural language dialogue system, which I think will completely change human’s behavior interacting with the Internet, hence it has the potential to change the entire landscape.
What activities do you think are important in the fields of digitization and connectivity?
Personally, my key focus as the Director of AI research center is on artificial intelligence, but intelligence is only one part of the entire digital transformation agenda. Specifically, I see artificial intelligence as neither the beginning nor the ending part of digital transformation, but rather as one – important – component of the end-to-end digital revolution. In this regard, the industry conducts multiple important activities that can be grouped along the value chain. First, connectivity plays a critical role to the world by enabling the transmission of data. Huawei’s technology is responsible for about one third of the connectivity worldwide. The next competing ground is not only connecting humans faster and better but also connecting goods. Second, data storage is an essential field which gains further importance at present, and a lot of players stride to play a significant role in that field. Third, data need to be leveraged to enable a more intelligent use – data without being processed is no use. This is where artificial intelligence plays a key role. Fourth, service activities become increasingly relevant to unleash the power of intelligence. AI and cloud will be a strong marriage for the future.
AI and cloud will be a strong marriage for the future.
What role does artificial intelligence (A.I.) play in these activities?
The biggest role that today’s AI technology will play in all the activities we talked about above is the new mindset. In early 2014, I visited a very famous UIUC computer vision professor, who spent half century on this domain and mentored over 100 PhD students. He told me that one of his 3rd year PhDs was able to beat his decades of research, by using deep learning technology. This PhD student, undoubtedly, doesn’t have more knowledge about either computer or vision than my friend, the professor, has, but the student uses the power of data and computational power to reach the level that a lot of domain experts have long been hoping for. When we change an angle to look at the same problem, we may reach a totally different and, sometimes, better answer. We are approaching a paradigm shift where “data programming” is going to replace “logical programming” in increasing number of places. Domain expertise is facing fierce challenge from the magic of data. This direction is being intensively explored now, and there is still a lot of great potential yet to be discovered.
I see today’s AI as a natural extension of big data. In particular, today’s advances in AI are primarily supervised learning which relies heavily on big volumes of labelled data trained in structurally complex model. There are various efforts in the top research institutes working on unsupervised learning, transfer learning and reinforcement learning, which are critical to a sustainable advancement of AI, but the reality remains, for now, towards heavy reliance on data and computation power. This is a strength and a weakness at the same time for today’s AI technology. “What works for it will eventually work against it”, this has special meaning here. Initially, today’s progress was made on the top of Moore’s law and large availability of Internet data. In the past, hundreds of thousands of human hours were spent on labeling one data set in order to train an image model. This approach is not scalable. Trainable data becomes the biggest bottleneck moving forward. The low hanging fruits are probably gone, we may have to work differently in order to continuously march forward.
Which further developments do you expect in the field of artificial intelligence, and what major challenges will have to be addressed?
The field of AI will move in different directions in the future. Many of these moves will be exploratory, and it is hard to tell where exactly the field will go in the future. Nonetheless, I see four important directions in the industry at present. Firstly, deep learning will continue to play a key role to arrive at better algorithms and improved models, with innovations in areas of memory and attention, for instance. Secondly, learning mechanisms will be a central issue in order to go beyond supervisor learning and to explore other ways for improving learning based on creating new learning mechanisms, primarily to reduce the dependency on labelled data and to improve the efficiency of utilizing computation power. Adversarial learning and recently release progressive learning are of particular interest to me. Thirdly, the robustness and reliability of AI will gain further importance because today’s technology has “inherited” weaknesses, and the reliability of learning has to be enhanced in the future, through system approach rather than algorithm approach, to my opinion. Fourthly, the computing architect will a bottleneck in the future. Therefore, innovation in chip-level and system-level is needed in order to overcome limitations with respect to enable larger-scale of parallelism.
Do you see any major differences between the artificial intelligence activities in different regions and countries worldwide?
Yes, different regions and countries certainly had and still have some particular emphasis in their AI activities. Consequently, collaborations at an international level should happen even more. Traditionally, many parts of Europe did not play a prominent role in AI research, but they have been improving significantly. Examples are important activities in Germany, Switzerland and the UK. Each area tends to have its own advantages. For instance, France and Russia have been very good in mathematics and algorithms. Other countries, such as Canada, also have particular strengths. For example, Japan and Korea have important activities related to AI for robotics, and they are extending these activities also to other fields. China has been very good in implementation and incremental enhancements, and it has recently extended its activities to more radical novelties. Finally, the US are famous for the start-ups in the AI field. They are certainly leading in terms of putting the technologies into business, with the most significant AI start-ups based in the US.
What are your own personal goals with respect to innovation and digitization in the next two years or so?
In a time-span of two years, I would expect that the natural language dialogue will have major breakthroughs in terms of language matching and dialogue modeling. There are some great and promising technologies that will enable natural conversation with AI. If we assume a more long-term perspective of five years, I see some focus on intelligent machines, namely automated driving and robotics, however driverless cars in generic terms, to my opinion, probably have too many challenges to be achievable before 2020. Narrowly-defined use scenarios are a much more plausible target for that time span. In intelligent machine domain, reinforced learning mechanisms may be especially helpful for achieving high reliability and robustness, which is key in the field of automated driving. Better computation vision beyond today’s object/scene classification poses the primary challenge in this domain, and system computation power, algorithm and sensory hardware definitely need further improvement. There are exciting times ahead of us with regard to AI.
Michael Wei is Director of Samsung AI research center, based in US. In this position, he is responsible for technology strategy, key projects and collaboration with universities and startups. Prior to joining Samsung in September 2016, Michael was Director of Huawei AI lab. He has 15 years expertise in intelligence technology with various positions from Lucent Bell Labs, IBM Watson, A.T.Kearney to Huawei. Michael received his MBA from UT @ Austin and Master in Computer Science from University of Southern California.
This interview was first published at Innoboard, a leading Innovation blog with a European perspective.