This post is also available in:
עברית (Hebrew)
As artificial intelligence systems grow increasingly complex, a new study from researchers at New York University proposes that the future of AI might benefit from a return to simplicity. Instead of immediately assigning AI systems high-level, multifaceted tasks, the researchers argue that training them gradually, similar to how young children are taught, could lead to more reliable, adaptable performance.
The study, published in Nature Machine Intelligence, introduces a framework where AI agents, like humans, are encouraged to master fundamental behaviors before being exposed to more intricate challenges. Drawing inspiration from early childhood education, the research emphasizes sequential task learning, reward-based motivation, and gradual complexity as critical components of quality AI training.
The initial phase of the research involved lab rats trained to perform a series of ordered actions in a controlled environment. Using sound and light cues, the rats were tasked with retrieving water from specific ports in a compartmentalized box. Success depended not only on recognizing the correct cues but also on understanding timing and reward delay. These simple, layered behaviors allowed researchers to monitor how well the rats learned to associate sequential steps with outcomes.
Translating this process to machine learning, the team applied similar principles to recurrent neural networks (RNNs)—an AI architecture used for processing sequential data. However, unlike humans, AI often struggles with memory retention when tasked with multiple objectives at once, leading to errors or inconsistencies in task execution.
To test their hypothesis, the team trained RNNs using a decision-making simulation that involved wagering on outcomes. The AI systems demonstrated the ability to solve simple tasks first, and then showed improved performance in more complex scenarios. The incremental training enabled better generalization and decision-making, critical abilities in real-world autonomous systems.
The researchers suggest adopting a “kindergarten curriculum” approach for AI: simple, clearly defined tasks that build toward higher-order capabilities. This method could enhance AI’s performance in real-world applications, where environments are unpredictable and multitasking is required.
As AI continues to influence sectors from defense to healthcare, this call for foundational learning serves as a timely reminder that even the most advanced systems may need to master the basics before aiming for sophistication. Just like in human development, the road to intelligence may well begin with learning to crawl before walking.
