An anaesthetized ferret watched a geometric image move across a computer screen. The shapes were designed to stimulate the parts of the ferret’s brain that decode lines, so that researchers could use a brand new technique to watch what its visual neurons responded to. What they found may indicate that we’ve underestimated the total computing power of carnivores’ brains. Or, at least, their observations might assist in refining the brain-mimicking software that Google used to win Go and IBM used to win Jeopardy.
Specifically, the researchers looked at neurons known for helping carnivore’s understand edges. Determining edges allows our brains to separate objects cognitively, so a creature knows which part of a scene is just a tree and which part is a jaguar preparing to pounce. More importantly for our purposes, though, is what they observed as data from a scene reached the neurons they were watching.
Neurons are the brain’s transistors. Computer processors pack thousands of transistors together. “Basically, we have spent a lot of time, trying to estimate the computing power of neurons,” Dr. Ben Scholl, one of the researchers at Jupiter’s Max Planck Florida Institute for Neuroscience told us. He was part of a team that investigated how neurons associated with vision deal with the enormous number of inputs that come into them. In research published recently in Nature Neuroscience.
Neurons either like their inputs and fire, or they don’t and do nothing. Like real computers, it’s the sum total of firings (or not firing) across thousands or millions of neurons that collectively make up each thought or image. Neurons have dendrites extending out from them, gathering up electrical signals. These have been mostly thought of as wiring up to now, but the Max Planck team observed the dendrites helping.
When a neuron has a decision to make, the brain ends up sending it a lot of irrelevant data. The dendrites studied by Dr. Scholl and his colleagues were lined with these spiney structures, though, and those structures were guiding the useful bits of data in such a way that made the neurons’ jobs easier.
Imagine the brain as a giant seaport filled with boats going to cities all over the Mediterranean, but only cats used the port. The cats here are the visual data. Their owners put hats on the cats’ heads, and the colors of the hats indicate which destination the cat should go to. A red hat, for example, would mean it was meant to go to Majorca (just as certain neurons only fire if the data it receives corresponds to vertical lines).
Once released onto the docks in the port, the cats (being cats) would run onto whatever boat struck their fancy. Which is pretty much what the data that comes into our brains appears to do.
So, in this story, the boats in port are the neurons. If the captain of the Majorca-bound boat comes out of his cabin and sees enough cats with red hats on, he will order his ship to launch. If not, though, he would order whatever cats had come on board thrown overboard and go back into his cabin to await the next deck full of cats.
The dendrites are the docks the cats are running around on. We might have once thought those docks were passive pathways to the boats, but they aren’t. There are workers on the docks herding the cats (these are the dendrites’ spines). The workers grab cats, who happen to be heading to the right boat and gather them up so they can go up the gangway in bunches. That way, when the captain comes out, bunches of the right color will be easy to spot.
So then when the Majorca-bound captain leaves his cabin, he can easily see the cats in red hats bunched together. So, he’ll order the boat to pull out for Majorca. (and the cats with the wrong hats who had also come aboard will just get thrown over)
It’s an admittedly mean metaphor.
What the research team saw in the ferret’s brain was that, while the dendrites’ spines aren’t filtering information for their neurons, the spines do appear to be organizing it. When information that a certain neuron will react to comes its way, the dendrites make sure those inputs get noticed.
So there’s more work taking place here than neuroscientists might have thought. Which could mean that our calculations of the brain’s total power are off.
“It’s processing,” Scholl said; however, more research will need to be done to see if these observations apply beyond visual processing.
“It’s totally possible that what we’re looking at is not specific to vision,” he granted. In other words, dendrites might organize inputs in similar ways as it processes sound or textures.
Neuroscientists have been mapping which parts of the brain analyze different kinds of data and how much work overall gets done in there. “I think that a lot of the work that we’re doing is we are trying to understand the basic building blocks of the way things are wired up,” Scholl said, but these insights apply outside neuroscience.
Our understanding of the brain eventually has real impacts on the computing infrastructure will power the most compelling applications of tomorrow. Artificial intelligences work on what are called deep neural networks, which are based on a rudimentary understanding of how the brain works first modeled decades ago, Scholl explained. As our model of the brain becomes more refined, engineers building artificial neural networks could sharpen their machines based on what neuroscientists discern.
So as researchers keep pursuing these questions of how brains deal with the overwhelming amount of information they constantly take in, Scholl noted, “we could maybe understand how the brain is working but also get a better understanding of how to train our deep neural networks.”
That’s when computers may begin surprising us. Artificial intelligence can already compose music, but it could then make really good music. A.I. can already recognize cats in pictures, but it might send drones out to compose and shoot really beautiful photos. In fact, we might not really understand how our own brains work until an artificial intelligence takes over and looks into it on our behalf.
