Infants naturally learn about the world through a process of incremental understanding, building foundational concepts before tackling more complex ones. For example, infants recognize animacy - distinguishing between living and non-living entities - before learning about goal-directed behavior. This stepwise learning allows them to make accurate predictions about future events, such as expecting a hand to grasp an object rather than move randomly.

The ability of infants to acquire and understand concepts with minimal supervision has long fascinated researchers. Unlike artificial intelligence (AI) models, which often require vast amounts of data and computational power, infants learn quickly by building on early-acquired concepts.

A recent study titled From Infants to AI: Incorporating Infant-like Learning in Models Boosts Efficiency and Generalization in Learning Social Prediction Tasks, authored by Shify Treger and Shimon Ullman from the Weizmann Institute of Science, explores how mimicking infant learning strategies can enhance AI’s performance. Published in arXiv (2025), this study demonstrates that integrating infant-like concept decomposition into AI training leads to improved learning efficiency, higher accuracy, and superior generalization to novel tasks. The findings highlight fundamental differences between human and AI learning and propose new methodologies to bridge the gap.

How infants learn and why AI struggles

Infants naturally learn about the world through a process of incremental understanding, building foundational concepts before tackling more complex ones. For example, infants recognize animacy - distinguishing between living and non-living entities - before learning about goal-directed behavior. This stepwise learning allows them to make accurate predictions about future events, such as expecting a hand to grasp an object rather than move randomly.

In contrast, traditional AI models rely on end-to-end learning, where all patterns and relationships are learned simultaneously without decomposition into simpler concepts. This approach, while powerful, has limitations. AI models often struggle with generalization, requiring massive datasets to perform well on new tasks. Additionally, they are prone to errors when faced with situations slightly different from their training data. The study proposes that incorporating human-like concept decomposition into AI models can significantly enhance learning outcomes, making them more efficient and adaptable.

A cognitive approach to AI learning

To test their hypothesis, the researchers compared two AI training methodologies: the Cognitive Model, which mimicked infant-like concept learning, and the Naïve Model, which followed the traditional end-to-end learning paradigm. The study focused on social prediction tasks, where models had to determine the future actions of animated and inanimate objects based on prior interactions. Using a custom-designed dataset, the Cognitive Model was trained to first distinguish between animate and inanimate objects before learning action prediction, while the Naïve Model learned both tasks simultaneously.

The results were striking. The Cognitive Model:

• Achieved higher accuracy in predicting future events compared to the Naïve Model.
• Required significantly less training data to reach peak performance.
• Demonstrated superior generalization, performing well even when exposed to completely new actors and scenarios.

By introducing a structured learning process, akin to how infants acquire knowledge, AI models became more efficient, reducing the need for extensive data while improving their adaptability to novel situations.

Implications for AI and Machine Learning

The findings of this study have profound implications for AI development, particularly in areas requiring human-like reasoning and adaptability. Current deep learning models, while effective in specific tasks, often fall short when asked to generalize beyond their training data. By integrating early concept learning into AI training, we can develop models that require less data, learn faster, and perform more robustly across varied environments...

Read the whole story on the Devdiscourse website using the link below.