Today’s most advanced AI models lag behind one-year-old babies in terms of learning efficiency, despite using thousands of high-performance chips and huge information sets. Researchers state that the main reason for this situation is that the learning architecture that the baby brain uses to make sense of the world is completely different from existing systems.

Babies can recognize a new object after seeing it only once or twice, and can quickly comprehend their surroundings with short periodic observations. On the other hand, artificial intelligence models need much more data and power to achieve similar feats.

Babies’ learning process serves as an example for artificial intelligence

Meta, Stanford University, Tokyo University and Ecole Normale Superieure researchers developed a new evaluation system called EgoBabyVLM Challenge to analyze this difference. In this system, artificial intelligence models that can process visual and textual information are asked to analyze approximately a thousand hours of real-life images recorded by cameras attached to babies’ heads.

The results show that even the most advanced models fail to perform as expected when faced with real-life complex landscapes rather than clear information sets. Experts say that cognitive systems that enable babies to learn quickly from small amounts of information offer valuable clues for developing more efficient artificial intelligence systems in the future.

Cognitive scientist Michael Frank from Stanford University emphasizes that the learning process of babies is not just about words. Babies listen to their parents talk about objects that are not in their field of vision, follow social cues such as gaze direction and hand movements, and make sense of the world by combining sensory experiences such as touch, sight and hearing.

Difficulties continue to make sense of the physical world

The BabyLM project, developed in 2023, gave successful results in learning the syntax of the language by training artificial intelligence with as much data as a child is exposed to. However, researchers state that the same success cannot be repeated in understanding the physical world.

Joshua Tenenbaum from MIT states that although current models are successful in capturing patterns in information, they cannot develop skills such as common sense, social interests and physical reasoning that a baby acquires. Researchers think unraveling how babies achieve the comprehensive cognitive feats they do by age two is a critical threshold for future AI technologies.

Stanford’s Michael Frank and his team are working on a new type of model that is better at learning causality and visual and temporal relevance. It is anticipated that this approach, which can grasp object dynamics more effectively, can contribute to the development of artificial intelligence systems that learn faster in fields such as physics and social affairs.

When do you think artificial intelligence systems can reach this natural learning and reasoning ability that babies have?