AI Week interview: reemergence of the central role of knowledge in AI's Third Wave
Amit Sheth
Founding Director, Artificial Intelligence Institute at University of South Carolina
This is a rough transcript of the interview I completed for the AI Week (Speakers). I will post the link to the video closer to the event (May 10-14, 2021). For additional related thoughts, see: "The Duality of Data and Knowledge Across the Three Waves of AI."
AI Week: Tell me about your personal journey with AI — you’ve been working in this field for a few decades, what have you learned and seen? Where is AI now?
My first job was with Honeywell Research in 1985, where I did prolog programming and built an expert system. My second job at Unisys research, where I led a 1.6 million dollar AI and DB integration project funded by DARPA around1988. It was about integrating interpreted to compiled logic-based systems with databases. The next major AI milestone in my life was Taalee/Semagix which I founded in 1999. It was the first commercial semantic search engine that used a large and dynamically updated ontology/KG to provide semantics to search—12 years ahead of Google’s knowledge graph reliant semantic search. We also got the first patent on this area that was titled: Semantic web and its applications in browsing, searching, profiling, personalization and advertising.
In 2005, we talked about the complementary nature of statistical AI/ML, symbolic AI with formal knowledge, and beyond: Semantics for the Semantic Web: the informal, the formal, and the powerful. This is highly relevant to the current interest in neuro-symbolic AI that is part of the third wave.
In terms of where AI is now, there is a good bit of hype but also a significant and growing number of real-world successes. The AI Institute at the University of South Carolina works with most colleges in our comprehensive university, on topics ranging from medicine, public health, life sciences, future manufacturing, neuroscience, astrophysics, education, and social good. This translational aspect is just one of many indicators of the massive impending impact of AI. The same is happening in Industry. One of my recent talks was directly based on what the CEO of Bain Capital had said at the World Economic Forum: “Every company now is an AI company. The industrial companies are changing, the supply chain…every single sector, it’s not only tech.” And at the same time, I have also said that we will not see singularity in my lifetime.
AI Week: One of the interesting conversations in AI right now is around the integration of data & knowledge and how it will fuel the next wave of AI. Tell me about your work on that.
Let us look at the analogy from human intelligence—how the human brain uses data to go to decisions. A human brain is bombarded with an estimated 11 million bits per second of sensory data, but our conscious brain processes the data into information, knowledge, and wisdom needed to make decisions and take actions. Knowledge plays a critical role. Cognitive Scientists have talked about the bottom brain and the top brain. Nobel prize-winning economist, Daniel Kahneman talks about system 1for thinking fast, for perception, and system 2, for thinking slow that involves planning and cognition. Neuroscientists have also recognized the dichotomy, but I expect future work to identify the explicit role of knowledge and experiences when converting data into decisions and actions. All this tells portends a significantly enhanced and accelerated role of knowledge to go beyond big data that deep learning has relied on in the second wave of AI.
AI Week: Taking that into a more specific angle, tell me about how this works at a more practical level. What does this third wave of AI mean for industries like healthcare, education and others?
In recent years, my team has focused on the theme we call Knowledge-infused Learning, which I hope will drive the progress in neuro-symbolic computing, or hybrid AI that seamlessly combines statistical AI and symbolic AI with knowledge playing the critical role in integrating the two. We have described a range of alternatives spanning shallow, semi-deep and deep infusion. The general idea is that in shall infusion, we map knowledge into the forms—called embeddings—that can be easily incorporated in current deep learning algorithms, albeit at the loss of deeper semantics. Going towards deep infusion, we explicitly recognize the role of abstractions and use stratified knowledge to align with the layers of implicit knowledge abstractions along with the layers in deep learning. In my mind, along with context and personalization, abstraction is critical in building AI systems that can effectively mesh into humans activities and can bring machine intelligence to become synergistic to human intelligence.
Statistical AI has done wonders for some industries, but healthcare and education demand more from AI systems. For example, it is important to incorporate medical knowledge and clinical guidelines—the way clinicians think, work and operate, in the AI systems. The use of knowledge is also critical in building an explainable AI system which is required in medicine. In most cases, a doctor cannot afford to rely on an AI system that inscrutably predicts some prediction or recommends some clinical diagnosis- it is necessary to explain a decision in terms of the clinical data or observations and settled clinical knowledge. Think of using a trellis to train a vine—the knowledge representation provides a way to pin the statistical patterns onto concepts that are recognizable to a human.
AI Week: How does all of this come together to set up the next wave of AI? What does that mean in practical applications for people in the future?
While statistical AI has surpassed human level of performance in certain well-defined tasks, such as computer vision, in the third wave, I see the AI system become more capable of integrating into human life, consistent with a vision I had described over a decade ago, and called it “Computing for Human Experience.” I have done four startups—three of them involve the general philosophy of knowledge-empowered AI. I also advise a couple of EduTech companies and a nutrition company—all of which are heavy users of knowledge graphs, deep learning and natural language processing/understanding. Each of these companies have close or distant relationships with healthcare and education. Health and education are domains where it has been difficulty to see significant positive impact of technology in the past. But now I am bullish about large scale use of AI in improving healthcare and education in less than a decade, and the impact of this on human life will be massive.
Enterprise analytics systems, processes, and business value
3 年Knowledge-powered AI and hybrid AI hold great promise. The "Narrow AI" term may also apply. An article that outlines the semantics of AI for business decisions: https://rahul-saxena.medium.com/narrow-ai-for-decision-intelligence-ada9e0d01001