Forget ripping off brains for AI. Butterflies and worms could lead us to self-repairing intelligent robots, says prof
'It’s clear that intelligent behavior doesn’t always require a brain'
The inner designs of today's artificial intelligence are often inspired by the human brain, yet there other biological structures perhaps better suited to crafting next-gen machine-learning software and hardware.
“Biology has been computing at many scales long before brains came on the scene,” Michael Levin, a professor and director of the Allen Discovery Centre at Tufts University in the US, said during a keynote speech at the NeurIPS (Neural Information Processing Systems) conference in Montréal this week.
He pointed to insects that undergo metamorphosis, and two-headed flatworms and salamanders that can regrow limbs. These processes are performed by a network of cells that work together to fulfill a specific goal, and could be inspiration for implementing intelligence in code as an alternative to the usual neuron-based approach.
In fact, some researchers have modeled a moth brain using software and trained it to recognize numbers. There is also another similar project, where neural networks were built to use information encoded as simulated DNA molecules.
The natural world displays signs of intelligence that have not been observed in machines. Levin compared a caterpillar to a soft robot. As it evolves, it sheds its skin to reveal a harder layer called the chrysalis. Inside, the caterpillar - including its brain - turns to mush and reemerges as a butterfly. Its brain has been reassembled to operate a new hard body that is able to fly around.
During the whole metamorphosis process, the memories of the caterpillar remain intact and are transferred to the butterfly. “Things that the caterpillar has learned, the butterfly still remembers,” Levin explained. “It’s very interesting to think about how this information can be stored in a medium that is undergoing radical deformation.”
Don't you forget about me
Today's artificial neural networks vaguely modeled on parts of the human brain lack this sort of plasticity, however. Bots that are trained to play a specific game, whether its Go, Super Mario, or Montezuma’s Revenge, cannot learn to play another game without forgetting how to play the first one. The problem known as “catastrophic forgetting” doesn’t seem to happen at the cellular level for some organisms.
Plenarians, a type of the jelly-like flatworm, can be chopped up into 256 pieces and each body part will seemingly magically regenerate into a completely new flatworm body. These animals also retain the same memory and knowledge as the original flatworm.
The ability to regenerate is also seen in salamanders, frogs, and even humans to an extent. It’s something that can be computationally modeled. Levin called it a “closed loop system”, where information, whether its how to transform caterpillars into butterflies or grow a new head, flows around a feedback loop until the task is completed.
AI threatens yet more jobs – now, lab rats: Animal testing could be on the way out, thanks to machine learningREAD MORE
In biological systems, the instructions are encoded in genes to make proteins that each carry out a specific bodily function. “Fundamentally, this is a computational problem. We’d like to use this as inspiration for technology that can implement this kind of highly robust, plastic behavior,” he said.
Levin’s lab has proved that biological changes, such as the shape of a flatworm, can be changed by manipulating its electrical synapses across the body. Its elongated body can suddenly turn spiky, or fold into itself like a bowl or top hat. All this can be done without explicit gene editing.
“Living things are the best computers. It’s clear that intelligent behavior doesn’t always require a brain, like single-celled organisms or slime molds,” Levin told The Register.
“We’re finally at a position, where we can explore these technologies for machine learning. For example, if we get away from mimicking the brain and take cues from living things, we might be able to build robots that can self-repair.” ®