Just as children learn to read by first mastering letters and perform arithmetic by understanding numbers, artificial intelligence benefits from a similar foundational approach, according to a new study.

Researchers,writing in Nature Machine Intelligence,have demonstrated that recurrent neural networks (RNNs) show improved performance on challenging tasks when initially trained on simpler cognitive exercises.

This method, dubbed “kindergarten curriculum learning” by the study’s authors, emphasizes the importance of establishing a solid understanding of basic tasks before tackling more intricate problems.

Cristina Savin, an associate professor at New York University’s (NYU) Center for Neural Science and Center for Data Science, explained the concept with an analogy: “From very early on in life, we develop a set of basic skills like maintaining balance or playing with a ball.”

“with experience, these basic skills can be combined to support complex behavior-as an example, juggling several balls while riding a bicycle,” Savin added.

“Our work adopts these same principles in enhancing the capabilities of RNNs, which first learn a series of easy tasks, store this knowledge, and then apply a combination of these learned tasks to successfully complete more sophisticated ones.”

RNNs,which excel at processing sequential data and leveraging stored knowledge,are commonly used in applications like speech recognition and language translation. However, conventional training methods often struggle to equip RNNs for complex cognitive tasks, failing to replicate key aspects of human and animal behavior.

To overcome these limitations, the research team, including David Hocker and Christine Constantinople, both of NYU’s Center for data science, began with experiments involving laboratory rats.

The rats were tasked with finding a water source within a multi-port box. To succeed, they had to associate water delivery with specific sounds and light cues, and also learn that water wasn’t instantly available after these cues. This required the rats to develop a basic understanding of multiple elements and combine them to achieve the goal of obtaining water.

These experiments revealed how the rats utilized their knowledge of simple tasks to accomplish more complex ones.

Inspired by these findings, the scientists trained RNNs using a similar approach. Instead of water retrieval, the RNNs were given a wagering task that demanded they build upon basic decision-making skills to maximize their earnings over time. The team then compared this “kindergarten curriculum learning” method against existing RNN training techniques.

The results indicated that rnns trained with the kindergarten model learned more efficiently than those trained using conventional methods.

“AI agents first need to go through kindergarten to later be able to better learn complex tasks,” observes Savin.

“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.”

“AI agents first need to go through kindergarten to later be able to better learn complex tasks,”