By Rob Mitchum // March 17, 2016
As interest in artificial intelligence gains mainstream traction, computers and brains are increasingly compared to each other. But despite great advances in computer architecture and software, even the world’s most powerful computers can’t compete with the efficiency of the human brain. Capable of processing a constant flood of sensory information and make snap decisions, the brain nevertheless only burns about 15 watts of energy, less than a common light bulb.
This unfair battle has two implications for science, said Narayanan “Bobby” Kasthuri in his talk at the Computation Institute on March 10th. On one side, computers struggle to comprehend the complexity of the brain, even as imaging techniques improve at capturing neural architecture. But in reverse, scientists are interested in what the wiring of the brain, also known as the “connectome,” can tell us about designing more powerful and efficient computers.
Kasthuri, the first neuroscientist hired by Argonne National Laboratory, is primarily interested in the first of those relationships: developing new computational methods to help map and unravel the brain’s anatomy. Kasthuri’s talk began with Santiago Ramon y Cajal, the anatomist who first observed and illustrated the brain’s primary network of neurons and synapses. But despite the incredible and influential advances of Ramon y Cajal’s work, “these early, simplified diagrams of nervous system are woefully inaccurate compared to the reality of the nervous system,” he said.
In Kasthuri’s work, he developed new techniques to capture the full detail of brains using different forms of microscopy. His work with electron microscopy and x-ray spectroscopy have created some of the most detailed (down to ~1 nanometer resolution) snapshots of the brain ever collected, with advanced computational techniques used to transform 2D images into elaborate 3D diagrams. However, the results of these mapping expeditions reveal anatomy so complicated as to be impenetrable to the naked human eye.
“Our capacity to collect data is not the rate-limiting step,” Kasthuri said. “But there are not enough human beings in the world to map even the mouse brain at scale, tracing individual components. It will have to be done by algorithmic machine vision and learning.”
With help from computational scientists, Kasthuri hopes to discover patterns and structure within the tangle of brain cells that help better explain the mechanism of the brain’s impressive abilities. In turn, that knowledge could help inform future design of computer architecture and programming to improve speed and reduce energy cost, much as simplified neural diagrams have inspired neural network and deep learning algorithms.