The seven sessions in the artificial intelligence/machine learning (AI/ML) track at Upgrade 2024 leaned toward practical implementations, with a few exceptions. In this article, we’ll summarize the first three sessions: a discussion of AI involving two NTT partners, a theoretical talk by a NTT Research Physics & Informatics (PHI) Lab scientist, and a look at how AI can reduce technical debt. A follow-up article will cover the remaining four.
The first of the seven sessions in this track highlighted innovation that NTT is nurturing through ecosystem partners. Specifically, it featured two companies that received early funding from NTT Venture Capital (VC): Private 5G network technology provider Celona and Eko Health, a pioneer in applying AI for early detection of heart and lung diseases. Eko CEO Connor Landgraf said they partnered with NTT for its data and industry expertise. “We wanted investors who really deeply understood the challenges and complexities of working with data at hospital system scale,” Landgraf said. Eko was founded in 2013 and now has “about half a dozen FDA clearances and about half a million providers around the world using our devices and software and AI in clinical care.”
Celona CEO Rajeev Shah listed several advantages of the NTTVC funding. NTT validated their product, became the leading channel to market, delivered quality customers, and provided valuable product feedback and roadmap definition. As for AI, Shah had a personal comment: “It’s actually the first time in my entire career here that I have been confronted with a technology that neither can I fully understand, nor can I fully grasp the potential.” On the plus side and in the near-term, AI is driving Celona prospects to take another look a private 5G networks. “I’m both excited and scared for the next decade, honestly,” he said.
In the next session, PHI Lab Scientist Hidenori Tanaka pointed out that AI, despite becoming increasingly ubiquitous, still appears “alien” because of our limited understanding. It also can miss things that are obvious to a human, like counting or spatial relationships; or make other errors, like exhibiting bias. But unlike traditional computing, which simply executes human-written algorithms, with AI the computer can become “smarter” and deliver sometimes surprising results. But how do these insights – in both animate beings and inanimate machines – actually emerge? Last year, PHI Lab Scientist Gautam Reddy delivered a paper quantifying “a-ha” moments in mice, and in his talk Tanaka shared graphs that showed AI reaching similar “breakthroughs” while performing arithmetic calculations, transliterating phonetic alphabets, and solving math word problems.
Tanaka then presented a conceptual graph for a theory of imagination (an idea he said would have been laughable five years ago) that shed light on how AI/ML learns by combining concepts. “We can build a hypothesis saying maybe when the machine learns something, the machine goes on some particular progression of memorizing what [it] has already seen,” he said. Extricating computational principles from actual neural activity is one of the goals for this emerging field of study called the “Physics of Intelligence.” In a conference paper last year that leaned on large datasets of recorded neural brain activity in mice, Tanaka et al. introduced a new training method that proved 100 times faster than traditional optimization approaches, while maintaining or improving modeling accuracy.
Despite such tremendous leaps forward, Tanaka sounded a cautionary note. “There’s no doubt that AI is transforming society, but the whole point and the mission of NTT research and this conference has been to upgrade reality,” he said. “We’re going to make sure that the direction is going up rather than [toward a] random, arbitrarily bad-world dystopian society.” To that end, he emphasized integrating all layers of society, from bridging several scientific fields to sharing the benefits of AI through corporate collaborations to interacting with public officials on questions of regulation and governance.
While scientists are still asking basic questions and even trying to understand how exactly some AI/ML works, practitioners in the field are deploying large language model (LLM) tools to solve pressing business problems. One is technical debt, lines of legacy code that are costly to remediate and expose organizations to cybersecurity risk and operational failures. Citing McKinsey, NTT DATA Managing Director, Banking Industry Partner, Madhu Magadi said technical debt in the financial services industry alone amounts to $1 trillion. “Unfortunately,” Mechanized AI CEO Charles Wright said, “every year we fall further in debt.” An NTT DATA partner in fintech, Mechanized AI has developed a code map technology that leverages several LLMs to automate the process of updating and summarizing code, allowing for faster, more efficient, and secure modernization projects. In the case of a standard legacy app consisting of 300,000 lines of code, the Mechanized AI platform achieved dramatic results. “We’ve taken a year off an entire project and over 22,500 man hours,” Wright said.
There is a big opportunity to help enterprises pay down their debt more quickly. First consider that the average modernization project is about 18 months to two years per application. “Just the analysis mapping and discovery portion … that we have discussed today consumes half of that,” Wright said. “If you can shave half the time off of a modernization project, and if you consider that McKinsey research has stated that there are banks out here with thousands of legacy applications, you can do the math at the value.”