A new test developed by researchers at Meta, Stanford University, the University of Tokyo, and France’s École Normale Supérieure aims to evaluate how well AI models can learn like babies. The EgoBabyVLM Challenge requires vision language models to describe the world after ingesting about a thousand hours of video collected from cameras on infants and toddlers. This test highlights the gap between current AI capabilities and the rapid learning observed in human infants. Source: wired
The test shows that cutting-edge AI models fail to interpret realistic and messy footage, unlike babies who learn efficiently from fleeting observations and physical interactions. Babies learn from a kaleidoscopic view of the world, including parents talking about objects not visible, using gestures, or discussing past or future events. This multimodal and tactile learning experience is essential for understanding the world, according to Michael Frank, a cognitive scientist at Stanford involved in the project. Source: wired
The EgoBabyVLM test is part of a broader effort to explore human intelligence through AI. A 2023 challenge called BabyLM tasked models with learning language syntax using data comparable to what a 10-year-old absorbs—tens of millions of words, versus trillions for AI models. This showed that transformer-based models can perform well, challenging some theories about hardwired language abilities. However, understanding the physical world remains a challenge for AI, as noted by Ryan Cotterell, a linguist at ETH Zurich. Source: wired
