CUPERTINO, Calif. — Researchers face fundamental mysteries understanding how the brain works, but their work promises breakthroughs in computing as well as health care. That’s the view of Phillip Alvelda, a serial entrepreneur whose latest work is at the intersection of neuroscience and electronics.
Alvelda helped organize a research program to create an implantable neural-electrical interface. More recently, he launched a startup with the ambitious goal of creating the digital equivalent of a hippocampus and cerebellum.
Researchers are now able to track signals in a million-and-a-half neurons, the entire cortex of a mouse, he reported. “We can put an image in front of a mouse and read out how its processed … to start to tease out the actual neural code,” he said in a keynote at last week’s Hot Chips event here.
“How information is coded in the brain is not known; maybe it’s not a code of signals and switches [like those used in today’s computers] but something based on the relative time of arrival of multiple signals in a shared channel,” he said, pointing to work on neural information theory that began around 2009.
Today’s deep-learning systems such as Amazon’s Alexa, IBM’s Watson, and Facebook’s convolutional neural networks are relatively siloed, “not able to generalize out of their domains.” “Our learning systems need a common code that’s relevant to sensory and memory integration,” he added.
Cortical.ai aims to create digital equivalents of a hippocampus and cerebellum. (Images: Hot Chips)
Others agree that today’s neural networks are relatively crude in comparison to the virtual GFlops of computing the brain manages at an estimated 30 W.
“The brain is built on a different computation model we only partially understand, and all this deep-learning stuff is heading in a different direction,” said Doug Burger, a veteran computer architect who helped design Microsoft’s recently announced Brainwave system for machine learning.
“We need breakthroughs to snap back to the biological model of computing or an investment in some other new model to find a new Moore’s law,” said Burger in an interview at Hot Chips. “The advantage of the biological model is that we know one exists, and we don’t know how much digital headroom there is ahead in deep neural networks.”
For his part, Alvelda believes the hippocampus serves as the brain’s integrator, “assembling sub-AIs into integrated meta-AIs.” He wants to build such a system at his startup, Cortical.ai.
The company, which is so new that it has not even established a headquarters yet, also aims to build a predictive system that mimics the brain’s cerebellum.
&ldquo ;In just the last couple of years, we have learned that the cerebellum has more neurons than the whole rest of the brain, so it’s not just used for motor control refinement,” he said. “The cerebellum is connected to the entire brain, and it is now believed to help project future states of cognitive processes,” such as knowing how to catch a ball.
Next page: Birth of a neuro-engineering industry