Training the computer to mimic human performance

Mapping the millions of miles of neuronal “wires” in the brain could help researchers understand how those neurons develop intelligence, personality, and memory. Sebastian Seung, professor at MIT, and his students have been building tools that will allow researchers to untangle some of those connections. To find connectomes, which are the mysterious connections that Seung and his students are after, they will need to employ vast computing power to process images of the brain.

Piecing together connectomes requires analyzing vast numbers of electron microscopic images of brain slices and tracing the tangled connections between neurons. It’s a meticulous process-each neuron takes hours to trace, and each must be traced by as many as 10 people in order to catch careless errors. Using this manual approach, finding the connectome of just one cubic millimeter of brain would take tens of thousands of work-years, said Viren Jain, who recently completed his PhD in Seung’s lab.

Jain and his research partners want to speed up the process considerably through high-powered computers. To do that, they are teaching the computers to analyze the brain slices, using a common computer science technique called “automated machine learning”, which allows computers to change their behavior in response to incoming new data.

With machine learning, the researchers teach computers to learn by example. They feed their computer electron micrographs as well as human tracings of these images. The computer then searches for an algorithm that allows it to imitate human performance.

“Instead of specifying the details of how the computer does something, you give it an example of what you want it to do and an algorithm that tries to figure out how to do what you want,” says Jain. After the computer is trained on the human tracings, it is applied to electron micrographs that have not been traced by humans. This new technique represents the first time that computers have been effectively taught to segment any kind of images, not just neurons.

The eventual goal is to use computers to process the bulk of the images needed to create connectomes, but the researchers expect that humans will still need to proofread the computers’ work.