According to MacLennan, that “back to the future” idea of using analog computing to understand the brain has returned to fashion after falling out of favor with early artificial intelligence researchers in the mid-1980s. That’s because modern researchers have recognized that, if we're going to achieve artificial intelligence comparable to what humans or even other mammals possess, we must first understand the human brain, which basically functions like an analog computer, he added.

“The analog processing of information is more efficient than digital processing of information. We’re so enchanted with the flexibility and speed of digital technology, but the tradeoffs are different,” MacLennan said. “Look at the brain, which uses components which are orders of magnitude slower than the transistors in our current technology, but yet it's still able to do things we can't do very well with our digital technology. Part of the reason is that it's using low precision analog computing, but in a massively parallel scale."



MacLennan cited Carver Mead, who was an innovator in VLSI (very large scale integration) digital circuitry, and his statement from the 1980s that the future of electronics is in analog VLSI. Mead based that conclusion on his studies, which indicated the brain is primarily an analog information processor, MacLennan said, and over the past 10 or 15 years, there has been increasing recognition of that and a subsequent rediscovery of the value of analog electronics. Much of that, he added, has been inspired by brain-oriented computing in general.

Part of the decline of analog computing, MacLennan said, was the desire for high precision computation, which was less expensive to attain using digital technology. However, the brain has shown us it doesn’t need high precision, and while individual neurons are inaccurate computing devices, the brain has parallel and huge numbers of neurons. With that, researchers have discovered that if you put enough of those low precision computing devices together, you can create a form of precise behavior exhibited by animals.

“For example, instead of representing the value of, say, the brightness of a particular place on your retina by a single neuron, it represents it by a whole population of neurons, each of which are kind of sloppy,” MacLennan said. “But when you look at the whole population together, you get a very accurate representation of what the illumination is at that point.”

Given that, MacLennan said there is still a great deal that we don’t know about how the brain processes information. While there are new discoveries every week, and researchers understand some principles, such as sensory systems, motor planning and memory, there are still huge gaps.

As he looks to the future of understanding neural processes, MacLennan sees two possible scenarios. The first, which is the continuation of the slow progress of the research, will likely parallel artificial intelligence research, though as researchers get ideas about how the brain might work, it can be tested in an AI system. Regardless of the results, those ideas can provide new avenues for neuroscience to pursue.

The second scenario, which MacLennan notes can’t be predicted, is that one big breakthrough. It could happen tomorrow or it could be several years down the line, but that discovery could provide neuroscience researchers with the basic principles that operate throughout the brain; these could then immediately be applied to our understanding of neural networks, he said.

Beyond those scenarios, MacLennan emphasized an open dialogue with interdisciplinary researchers is vital to furthering the study of neural information processing. The example he gives is the work done by psychologists and philosophers on the importance of embodiment, which shows the importance of robotics as a vehicle for investigating artificial intelligence. That work, MacLennan added, points to the importance of neural information processing in the future of robotics.

“We need robots that have the physical competence comparable to animals and we'd like them in small packages. We can see lots of applications for robots of various sizes,” MacLennan said. “This is why we need to investigate analog electronics. We're facing limits of digital technology and way one way around that is to get more computational use out of each electrical component. Analog computation is one way of doing that.”

Image: Dr. Bruce MacLennan