Making AI Humble? How Brain Development Teaches Models to Say "I Don’t Know"

This is a paper discussion event, requested by one of our community members. Why should you care about academic papers on AI? It usually pre-dates commercialisation by 5 years or longer. It's a sneak peak into the future, or at least a possible version of it.

Paper Link: arXiv:2412.17411 (Also published in Nature Machine Intelligence)

Can't join us in Canberra? Here's live event link:
https://youtube.com/live/iNsVG4CtgeA?feature=share

Have you ever interacted with an AI that was completely, confidently wrong? Whether it is a chatbot hallucinating a fact or a computer vision system misidentifying an object with 99% confidence, modern neural networks suffer from a major design flaw: uncertainty miscalibration. They lack the "meta-cognition" to know when they are guessing.
At this meetup, we will discuss a paper published in Nature Machine Intelligence (and originally hosted on arXiv): "Brain-inspired warm-up training with random noise for uncertainty calibration" (arXiv:2412.17411) by researchers Jeonghwan Cheon and Se-Bum Paik.
The authors made a surprising discovery: the standard practice of "randomly initialising" neural networks actually bakes in an overconfident bias before the model ever sees a single real data point. To solve this, they took inspiration from developmental neuroscience. Just as a mammalian brain experiences spontaneous, random neural activity in the womb before birth (such as prenatal retinal waves), they "warmed up" neural networks by pretraining them on purely random noise and random labels.
The result is a network whose confidence naturally aligns with its actual accuracy. When presented with completely unfamiliar or out-of-distribution data, it learns to lower its confidence and essentially say, "I don't know."

If you're joining on your laptop, phone, or via your AI assistant, here's the live event link:
https://youtube.com/live/BFYPAwpoDnc?feature=share