Artificial intelligence has seen significant advances in mastering language, diagnosing diseases, and solving protein structures, yet everyday tasks like dishwashing by robots remain elusive. This mystery may be resolved by studying the brain, according to neuroscientist Anthony Zador.
Zador, a professor of biology at Cold Spring Harbor Laboratory and noted expert in neuroscience and AI, delivered insights at the annual Swartz Foundation Mind Brain Lecture. In his speech titled "A Neuroscientist’s Guide to Artificial Intelligence," he discussed how understanding the brain’s computations could drive AI innovations, while AI might help unravel the brain’s mysteries.
“My lab actually does both neurobiology research — with actual organisms — and computational work in AI,” Zador explained. “But today, I’m going to focus on the AI side.”
Hosted by the Department of Neurobiology and Behavior at Stony Brook University and supported by the Swartz Foundation, the lecture aimed to bring advanced discussions on the brain to a wide audience. Zador acknowledged that despite advances, reality has not caught up with sci-fi depictions of intelligent machines.
“When I was a kid, I watched a lot of sci-fi,” he said. “And it was just assumed that in the not-too-distant future, we’d have intelligent AI controlling robots. Well, we’re now well beyond some of those timelines — and we still aren’t there.”
While AI systems have advanced, they still make unexpected mistakes. Large language models like ChatGPT can convincingly replicate human interaction but also make bizarre errors. AI vision systems can misidentify objects based on context, revealing AI's failure to generalize like humans.
“These kinds of errors reveal something fundamental,” Zador explained. “AI doesn’t generalize the way humans do. It gives more weight to patterns in the training data than to deeper, more flexible reasoning.”
Zador discussed the problem's roots, highlighting Moravec’s Paradox which indicates that tasks easy for humans are challenging for machines. Basic tasks like perception and movement are results of sophisticated brain computations that AI struggles to replicate.
He argued against the effectiveness of scaling up data and algorithms without limit. “Just because airplanes improved over the past century doesn’t mean they’ll ever fly to the moon,” he said. “There are fundamental constraints to how far brute-force scaling can take us.”
Zador proposes that neuroscience might offer solutions through understanding how the brain learns via structured world representations. This could help AI overcome its current limitations.
Zador is a leader in NeuroAI, aiming to merge neuroscience insights with AI research. This interplay has already led to breakthroughs, with neuroscience-inspired models advancing natural language processing. AI aids neuroscientists in mapping brain circuits with greater detail than ever before.
He suggests that the collaboration between AI and neuroscience could transform both fields and deepen our understanding of intelligence itself. “Ultimately, the goal is not just to build smarter AI,” he said. “It’s to unlock the secrets of the mind.”
---