Over the past few years, the development of advanced machine learning algorithms has enabled artificial intelligence to achieve remarkable performance in performing tasks and solving complex problems. However, machines still have great difficulty performing basic tasks of daily life. For neuroscientist Anthony Zador, in order to solve this problem, we must rethink artificial intelligence in terms of neural networks based on models of animal brains.
Artificial intelligence (AI) still has a lot to learn from animal brains, says Anthony Zador, neuroscientist at Cold Spring Harbor Laboratory (CSHL). He hopes that lessons learned from neuroscience can help the next generation of artificial intelligence overcome particularly difficult obstacles.
Zador has devoted his entire career to describing, from the single individual neuron, the complex neural networks that make up a living brain. But he started his career studying artificial neural networks (ANN). ANNs, which are the computer systems behind the recent AI revolution, are inspired by the branching networks of neurons in the brains of animals, including humans.
In an article published in the journal Nature Communications , Zador describes how improved learning algorithms allow artificial intelligence systems to achieve superhuman performance for a growing number of more complex problems such as chess and poker. Yet machines still struggle to perform what we consider to be the simplest tasks of daily life.
Resolving this paradox may finally allow robots to learn to do something as natural as stalking prey or building a nest, or even something as humane and mundane as washing dishes, a task which Google CEO Eric Schmidt once called "basic demand...but an extraordinarily difficult problem for a robot .
“Things we find difficult, like abstract thinking or playing chess, aren’t actually difficult for machines. The things we find easy, like interacting with the physical world, that's what's hard for them. The reason we think it's easy is that we've had half a billion years of evolution that has organized our neural circuits so we can do it effortlessly explains Zador.
That's why Zador writes that the secret to fast learning may not be a perfected general learning algorithm. Instead, he suggests that biological neural networks sculpted by evolution provide a kind of scaffolding that facilitates rapid and simple learning for specific types of tasks, usually those essential for survival.
>On the same subject:Artificial intelligence:it is on its way to changing scientific research forever
"You have squirrels that can jump from tree to tree a few weeks after birth, but we don't have mice that learn the same thing. Why ? Because the squirrel is genetically predetermined to become an arboreal creature, while the mouse is not he declares.
Zador suggests that one result of this genetic predisposition is the innate circuitry that helps guide an animal's early learning. However, these scaffolding networks are much less generalized than the machine learning that most AI experts seek. If ANNs identify and adapt similar circuitry, Zador argues, food processors of the future could one day surprise us with clean dishes.
