A new study by researchers from MIT and Empirical Health used 3 million person-days of Apple Watch data to develop a foundation model that predicts medical conditions with impressive accuracy. Here are the details.
A bit of background
While Yann LeCun was still Meta’s Chief AI Scientist, he proposed the Joint-Embedding Predictive Architecture, or JEPA, which essentially teaches an AI to infer the meaning of missing data rather than the data itself.
In other words, when dealing with gaps in data, the model learns to predict what the missing parts represent, rather than trying to guess and reconstruct their precise values.
For an image, for instance, where some portions are masked and others are visible, JEPA would embed both the visible and masked regions into a shared space (hence, Joint-Embedding) and have the model infer the masked region’s representation from the visible context, rather than the exact contents that were hidden.
Here’s how Meta put it when the company released a model called I-JEPA in 2023:
Last year, Meta’s Chief AI Scientist Yann LeCun proposed a new architecture intended to overcome key limitations of even the most advanced AI systems today. His vision is to create machines that can learn internal models of how the world works so that they can learn much more quickly, plan how to accomplish complex tasks, and readily adapt to unfamiliar situations.
Since LeCun’s original JEPA study was published, this architecture has become the foundation for a field that has been exploring “world models,” which is a departure from the token-prediction focus of LLMs and GPT-based systems.
In fact, LeCun even left Meta recently to start a company focused entirely on world models, which he argues are the real path to AGI.
So, 3 million days of Apple Watch data?
... continue reading