Technology & Innovation

To brain or not to brain

To brain or not to brain

Artificial intelligence researchers haven’t always looked to the brain for inspiration. That is changing, although many experts still focus on purely mechanical approaches

It may be surprising in retrospect, but the pioneers of artificial intelligence did not look to the brain for inspiration.

When the field of artificial intelligence (AI) was created at Dartmouth College in 1956, the study of psychology was dominated by behaviourists who saw the brain as an impenetrable black box; something that was beyond observation and therefore not worthy of study. Instead, they examined how inputs (stimuli) relate to outputs (responses).

Early experts in AI took the same incurious approach to the brain.

“AI was really about intelligence, not the brain,” says Tomaso Poggio, director of the Center for Biological and Computational Learning at MIT. “The connection with the brain is because the only definition of intelligence is the Turing test. People took that test as the challenge—to have a black box that behaved like a human.”

Early AI systems were written in programming languages, such as Lisp or Prolog, which bore no relation to how the brain computes. Even early artificial neural networks—ostensibly modelled on the way brain cells work together—turned out to be barking up the wrong tree.

But things began to change in the 1980s with the arrival of a new type of artificial neural network called “convolutional neural networks”. These systems were based on the work of David Hubel and Torsten Wiesel, the duo who won the Nobel Prize in physiology for discoveries about how the visual cortex works.

Today, these types of visual-cortex-inspired neural networks (also called deep learning neural networks) are creating a lot of buzz and a lot of money for some people. Google recently bought London-based deep learning start-up DeepMind for £400m. Netflix, Microsoft and Facebook are also big investors in deep learning neural networks.

"There is a race on right now between pure artificial intelligence people who are really not interested in the brain, and neuroscientists who are really only interested in the brain"

Neuroscience has not just inspired new kinds of software, but hardware too.

In a recent interview with New Scientist magazine, Dharmendra Modha, founder of the Cognitive Computing Group at IBM, said that after studying the brain for more than 125 years, computer scientists have “enough glimmers of insight into anatomy, physiology and behaviour” that they can start to build systems that draw on this knowledge. This is exactly what Modha has done in developing the TrueNorth chip, a chip that runs on artificial neurons and synapses, and communicates using action potentials rather than packets of data. Modha admits that the TrueNorth chip is based on a “cartoon” of how the brain works, but he has demonstrated that it can perform useful computations, such as pattern recognition, with very low power consumption.

Professor Poggio is certain that neuroscience will continue to help engineers build better systems. But, at the same time that artificial intelligence engineers are looking to the brain for inspiration, others are moving in the opposite direction; becoming more and more reliant on statistical techniques. IBM’s Watson is an example of the latter.

“There is a race on right now between pure artificial intelligence people who are really not interested in the brain, and neuroscientists who are really only interested in the brain,” says Professor Poggio. “The question is: who’s going to understand intelligence first? Is the neuroscientist going to understand intelligence because they understand how the brain produces intelligence, or are the engineers going to understand intelligence because they will be able to make an intelligent machine?”

Is it essential to understand the human brain to create true artificial intelligence? Share your throughs on the Future Realities LinkedIn group, sponsored by Dassault Systemes.  

Enjoy in-depth insights and expert analysis - subscribe to our Perspectives newsletter, delivered every week