AI is one of the main topics of this session of the InnoVEX 2018 forum. The first panel of the AI forum is titled “AI is the Driver for Innovation”. The panel featured distinguished speakers from Intel, IBM, Microsoft, NVIDIA, and Appier with Dr. Li Chen Fu of NTU as the moderator.
Intel: AI is More than just Algorithm
The panel started with Dr. Hanlin Tang, the Principal Engineer of Artificial Intelligence Products Group of Intel Corporation. He shared his main evaluation points for startups in the AI field; all of which are related to the data they have. His concerns are for the amount of data they have, where they got it from, and their data acquisition methods. In his opinion, having a strong data acquisition strategy is central for pushing forward success in AI.
Most AI startups today have trained a form of neural network algorithms which are deployed to enhance products in their particular services. This can be in the form of recommendation systems, data centers, edge analytics, video analytics, image recognition, speech recognition, or more. However, to take AI and startups to the next level, data acquisition strategies are required.
There are some challenges in the industry when trying to apply AI, particularly when moving from proof of concept to scale out deployment in enterprise. The present core of AI is deep learning, which scales much more with data than machine learning. The problem is building proof of concept from a small amount of data cannot reach a desired accuracy level. This can fail to convince the broader enterprise to adopt the proof of concept into production without realizing the scaling out of performance with the amount of data. To solve this, the enterprise will need to invest in developing datasets needed to achieve accuracy necessary for enterprise development.
Microsoft: AI's Main Goal is to Improve Efficiency
Michael J. Chang, the Director of Artificial Intelligence Research & Development Center of Microsoft said that people often talk about AI as a technology, but less about its applications. AI of today is reminiscent of children aged 4 to 6; it can focus and do something well, but only if it's taught how to; otherwise it cannot perform well. The main goal of applying AI in enterprise & corporations is to improve efficiency including in the manufacturing process, business process, profit margin, and lower cost.
AI as a technology evolves from its core technologies; image recognition, speech recognition, and natural language interaction. From there, it can perceive, learn, and reason to extend the capabilities of people and organizations; help people make decisions and be more efficient.
AI innovations in the future will need plenty of enterprise researching on application studies. To increase tech adoption, it is vital for companies to find the right problems to solve with AI. With the right amount of data and right technology application, AI can solve a lot of problems and create better lives for everyone.
NVIDIA: AI is to Enhance Humans
Prof. Simon See, Senior Director of Solution Architecture & Engineering of NVIDIA implied that the purpose of AI is not to replace humans, but to increase their efficiency and enhance the human performance. In support of AI development, NVIDIA has started Technology Centers in Asia Pacific with many universities. The purpose is to help researchers work in the country and collaborate with people in other countries. NVIDIA is also cooperating with Taiwan's Ministry of Science and Technology (MOST) in the Taiwan AI Initiative to cooperate with startups, and research communities. To help training in AI development, they also launched the Deep Learning Institute (DLI) to provide a source suite of curriculum for people to train in AI.
IBM: AI Offers Opportunities in Innovation
IBM has a long history in the field of AI. Mars Hsu, Business Executive of IBM Cloud stated that IBM is more aggressive than ever in AI. Currently, the available AI solutions are Narrative AI; similar to machine learning and simply able to identify patterns. Currently many companies utilize narrow AI for image identification using reinforced learning. The goal for the far future is general AI which can do almost anything. There are still so many unknown factors surrounding it. How it runs, how it obtains data, or how it analyzes information are all unknown. However, general AI will be able to build its own data, model, reason, and answer anything.
The AI of today is the so-called Broad AI, which is based on multimodal data containing compressed data and processes. This can include data from any industry regardless of their connection. Mr. Hsu believes this is where the value will be and it will grow in the next few years.
AI is also an exciting field to be in as it offers a lot of opportunity in innovation. Using AI to understand and interpret data will significantly propel innovation. It is also an important field, but the platform designed need a significant consideration. Mr. Hsu gave 3 points for consideration, they are: reimagine workflows, learn more with less data, and protect your insights.
“Reimagining workflows” essentially means that the right AI platform is designed with deep vertical expertise and enterprise level performance, including transparency, compliance, and security levels. To “learn more with less data” means that duplication or redundant data should be avoided whenever possible. Transfer learning should also be leveraged when possible to rebuild the learning model. “Protecting insights” include maintaining complete control over user's data, models, and intellectual property.
Appier: AI Puts Data into Action
From the startup side, Chih Han Yu, the CEO and Co-Founder of Appier stated that every company needs an AI strategy now. There are some similarities between AI today and internet in 1995 or mobile apps in 2005; but the entry barrier for AI is generally higher because it will need a significant amount of talents, computations, and data.
AI can put data into action by unifying data sets, expanding reach, building prediction models, and make the right content or context across different channels. The question then is can AI be as creative as a human? Mr. Yu believes that they can serve as a benchmark and in a certain way be as creative as a human. In the next 5 years, he expected to see creative AI which can be writers, explainable AI which can explain its decisions, and interactive AI which functions as a UI.
After the panelists were done with their speech, the audience members were invited for a Q&A session. The questions and answers from the panelists are listed below.
Question: With so Much Data Being Generated with IoT Devices, Which Predictive Models will be the Most Relevant for AI?
According to Dr. Tang from Intel, the generated data will still need to be manually labeled by humans to be usable to train the AI models. For him, this is a big limiting factor to see which predictive models will be the most relevant. He sees video as one of the key components where computer vision methods have reached a level of maturity where they can do a variety of tracking. This is where a lot of effort was spend to build hardware & software to enable both startups and companies to best take advantage of video predictive analytics.
Dr. See of NVIDIA used the example of factories which have been equipped with a large amount of sensors. These sensors will generate a large amount of data and will only be useful with trained and labeled data. One of the ways to use the data is abnormality detections where the systems in place and data are used to predict where the factories might encounter failures. They are also working with Baidu in China for traffic optimization utilizing traffic data. In short, the model will have many applications. The important questions then are can the right data be acquired and can the model be trained for the desired purpose.
For Mr. Chang of Microsoft, the current systems generally do not have a continuous learning loop. The labeled data which has been trained in the cloud or a server was then put in an untrained machine and used locally. The large amount of generated data is a very important problem for the industry to solve, but has yet to be solved. For the AI to work as intended, it needs to have a secure & fail-safe environment where data can travel, be processed, enter the model, train, and then redeployed. This problem is one that needs to be solved first.
Mr. Yu of Appier believes with such a heterogeneous way data is being stored and processed, there must be a way to unify the data or unify the application. This way, they can leverage each other’s outputs and then combined into a predictive model. The industries then will need to be considered from the point of data readiness and people’s readiness to accept the transformation.
Question: How Taiwan Can Get on the AI and Keep up with the Rest of the World?
Mr. Chang of Microsoft said that in terms of AI itself, Taiwan is not late. All the academic researches on the core technologies of AI already exist. The problem is that the research results cannot be connected to industrial applications. In the last year or half, society has been educated on the importance of AI. The AI adoption is happening; people just need to be patient for companies to apply the core technology into applications they’ve been using and their core business.
Mr. Hsu of IBM offered 2 points that needs to be considered. The first one is how government should pay attention to the emerging technology because a lot of information comes not from universities, but from open platforms. The second is how all good vendors already have good platforms for AI education where free materials are offered to invite people into the ecosystem.
Dr. See of NVIDIA believes a way to commercialize research results is necessary. The government would also need to focus in certain industry and application of AI to make sure there are no dilution effect which will minimize the effect of AI adoption.
Dr. Tang of Intel said that countries and tool builders, not just tool users can have effective AI organizations. He encouraged Taiwan to work with people to build the software and hardware tools necessary to build AI. Currently, the stack is so vertically integrated that those who can go beneath the tools today can get the most out of advancing AI anywhere.
Mr. Yu of Appier believes that idea and speed of execution are vital. The main disadvantage of Taiwan is the small internal market size which makes it difficult to achieve economy of scale to be a successful business. For this, it is important for Taiwanese companies to be able to grow their businesses overseas.
Watch the panel here: https://www.youtube.com/watch?v=whn1eAEbXlI
InnoVEX will Return in May 2019
InnoVEX 2018 hosted 388 startups, 17,867 visitors, and over 200 global investors; many of whom gave a positive impression of InnoVEX. After a stellar 2018, InnoVEX 2019 will be held in late May. As there will be a new exhibition hall for COMPUTEX, the scale of InnoVEX is also expected to grow even larger than this year. The registration of startups and speakers will be announced in October 2018.