AI, machine learning hampered by data accuracy: expert

Talk to a tech expert and you’re likely to get an opinion on AI—but not necessarily a positive one.

Hype around the tech’s use cases has sparked somewhat of a backlash from experts. You can count Gurucul ​​VP of Product Marketing and Solutions Sanjay Raja as a skeptic. He told IT Brew he sees the issue with the promise of the tech—which he refers to as ML/AI to include machine learning—as, in part, the reliability of the input data it builds on.

“The data that you look at and are provided is going to be critical to how you train the AI to get better,” Raja said. “But it’s also where the accuracy of that data has to be very precise and very important as well.”

Model citizens. Raja added that, due to the uncertainty about whether AI models are trained or simply rule-based, there’s a conflict developing between vendors and investors. Ultimately, it comes down to whether or not the promise lines up with the actual potential of the technology.

“The question is, how real is it? How effective is it, and what’s the application for it? And does that application make sense?” Raja said. “That’s where a lot of the skepticism comes from: Do they have something real underneath that?”

For more bullish experts like Wedbush Securities Managing Director of Equity Research Dan Ives, the promise of AI is the beginning of a “fourth Industrial Revolution” that is “playing out, front and center” for the industry—though he cautioned that full adoption is still a little way out.

“What we’re seeing now is the first wave of AI spending play out, from Nvidia to Microsoft to [semiconductors] to software,” Ives told IT Brew recently. “But on the consumer side, we’re not going to see full AI use cases till 2025.”

Potential, potentially. Raja sees AI as an effective tool that can be used to detect vulnerabilities, fill in gaps in human engineering, and generally automate a lot of processes that currently require manual programming. But the potential is still only as great as what it can learn from.

“AI still needs to be trained heavily to be effective,” Raja said. “And I don’t know if very many of the solutions are trained that well.”

One hurdle is model collapse—the possibility that AI will begin to be degraded from polluted data generated by other AI platforms. Last year, researchers at Cornell University found that “use of model-generated content in training causes irreversible defects in the resulting models, where tails of the original content distribution disappear,” a problem that “has to be taken seriously if we are to sustain the benefits of training from large-scale data scraped from the web.”