Science

Digitized worm brain learns a new trick

View 3 Images
C. elegans is a worm whose simple brain has been digitized as a basic neural network, and now that system has been taught a new trick
TU Wien
C. elegans is a worm whose simple brain has been digitized as a basic neural network, and now that system has been taught a new trick
TU Wien
The TU Wien researchers who worked on the project, from left: Mathias Lechner, Ramin Hasani and Radu Grosu
TU Wien
Making use of C. elegans' hard-wired instinct to move away from touch, the researchers trained the virtual worm to balance a vertical pole on its tail
TU Wien
View gallery - 3 images

The human brain is an absolute beast of a computer, running on the processing power of 100 billion neurons. Emulating that system could supercharge our supercomputers, but before we can hit those heights we'll need to start with something a little more simple. And they don't come simpler than C. elegans, a worm whose basic brain of only 302 neurons has been digitized. Now researchers have taught this virtual worm a new trick, without writing a single line of code.

C. elegans may not be using that brain to ponder the mysteries of the universe, but it's more than enough for the creature to get around, eat and keep itself safe from danger. Like most simple organisms, it runs mostly on hard-wired instinct. If you touch it, for example, the worm will wriggle away without having to think about it.

With its manageable number of neurons, C. elegans is the only creature to have had its neural system completely simulated as a computer program. Remarkably, the virtual worm retains all of its instincts, so a virtual "touch" can trigger the same kind of response, without a human coder telling the program what to do.

The new study, conducted by researchers at TU Wien, set out to determine whether the simulated C. elegans could be taught to perform a new task that no real worm would ever encounter – balancing a pole on the tip of its tail. Crucially, the team wanted to do so without writing any additional code.

Making use of C. elegans' hard-wired instinct to move away from touch, the researchers trained the virtual worm to balance a vertical pole on its tail
TU Wien

The researchers took advantage of the worm's reflexive response to touch. To keep the pole upright, the virtual worm had to move itself in response to the stimuli of the pole tilting. This was done by adjusting the strength of certain synapses, which is a key part of other machine learning algorithms.

"With the help of reinforcement learning, a method also known as 'learning based on experiment and reward,' the artificial reflex network was trained and optimized on the computer," says Mathias Lechner, co-author of the new study.

In the end, the computerized C. elegans managed to pull off the stunt. That bodes well for researchers looking to use biological neural networks as models of artificial intelligence.

"The result is a controller, which can solve a standard technology problem – stabilizing a pole, balanced on its tip," says Radu Grosu, co-author of the study. "But no human being has written even one line of code for this controller, it just emerged by training a biological nerve system."

The next step for the researchers is to experiment further to see what other new tricks it could teach the system.

The research was published online, and the team demonstrates their virtual worm in the video below.

Source: TU Wien

View gallery - 3 images
  • Facebook
  • Twitter
  • Flipboard
  • LinkedIn
1 comment
fb36
IMHO all biological brains could be room-temperature quantum computers where neurons working like qubits or registers of qubits. How we think? Any given moment our brains make a selection among an astronomical number of different things to think about, just like how quantum computers work (which makes them superior to classical computers; their ability to evaluate all possible selections in an instant)!