Despite the impressive performance of modern AI models, they still struggle to match the learning abilities of young children. Now, researchers have shown that teaching models like kindergartners can boost their skills.
Neural networks are typically trained by feeding them vast amounts of data in one go and then using this data to draw statistical patterns that guide the model’s behavior. But that’s very different from the way humans and animals learn, which typically involves gradually picking up new skills over the course of a lifetime and combining that knowledge to solve new problems.
Researchers from New York University have now tried to instill this kind of learning process in AI through a process they dub “kindergarten curriculum learning.”’ In a paper in Nature Machine Intelligence, they showed that the approach led to the model learning considerably faster than when using existing approaches.
“AI agents first need to go through kindergarten to later be able to better learn complex tasks,” Cristina Savin, an associate professor at NYU who led the research, said in a press release. “These results point to ways to improve learning in AI systems and call for developing a more holistic understanding of how past experiences influence learning of new skills.”
The team’s inspiration came from efforts to reproduce cognitive behavior in AI. Researchers frequently use models called recurrent neural networks to try and mimic the patterns of brain activity in animals and test out hypotheses about how these are connected to behavior.
But for more complex tasks these approaches can quickly fail, so the team decided to mirror the way animals learn. Their new approach breaks problems down into smaller tasks that need to be combined to reach the desired goal.
They trained the model on these simpler tasks, one after the other, gradually increasing the complexity and allowing the model to build on the skills it had previously acquired. Once the model had been pretrained on these simpler tasks, the researchers then trained it on the full task.
In the paper, the team tested the approach on a simplified digital version of a wagering task that mimics a real-world test given to thirsty rats. The animals are given audio cues denoting the size of a water reward. They must then decide whether to wait for an unpredictable amount of time or give up on the reward and try again.
To solve the challenge, the model has to judge the size of the reward, keep track of time, and figure out the average reward gained by waiting. The team first trained the model on each of these skills individually and then trained it to predict the optimal behavior on the full task.
They found that models trained this way not only learned faster than conventional approaches but also mimicked the strategies used by animals on the same task. Interestingly, the patterns of activity in the neural networks also mimicked the slow dynamics seen in animals that make it possible to retain information over long periods to solve this kind of time-dependent task.
The researchers say the approach could help better model animal behavior and deepen our understanding of the processes that underpin learning. But it could also be a promising way to training machines to tackle complex tasks that require long-term planning.
While the methods have so far only been tested on relatively small models and simple tasks, the idea of teaching AI the same way we would a child has some pedigree. It may not be long before our digital assistants get sent to school just like us.
Source link
#Teaching #Kindergartner #Smarter
