New paper derives exact uncertainty formula for wide neural networks
Replacing costly deep ensembles with a PDE-based shortcut for prediction variance.
A team of mathematicians and computer scientists from LSAF, LMBP, MaIAGE, and UCA has published a paper on arXiv (2606.05982) that cracks the problem of quantifying uncertainty in wide two-layer neural networks without the brute-force overhead of deep ensembles. Their key insight: instead of training many networks to estimate prediction variance, they directly solve a linear stochastic evolution equation derived from a trajectorial central limit theorem.
The team proves that the limiting fluctuation process around the mean-field limit is a centered Gaussian process in the dual of a weighted Sobolev space. They derive a closed-form covariance representation via a backward transport equation with a nonlocal source term, driven by the mean-field trajectory. By testing against the activation function at a fixed input, they obtain an expression for the limiting variance of network outputs. Numerical experiments on a 1D regression task confirm the approach works. This work could drastically reduce compute costs for uncertainty-aware AI systems in production.
- Replaces deep ensemble uncertainty estimation (training many models) with a single PDE solve, cutting compute by orders of magnitude.
- Proves the limiting fluctuation process is a centered Gaussian process in a dual weighted Sobolev space.
- Closed-form covariance derived from a backward transport equation with nonlocal source term, validated on 1D regression.
Why It Matters
Enables cheap, mathematically rigorous uncertainty quantification for wide neural nets, critical for safe AI deployment.