AI Model Reduces Parameters by 99 % While Keeping Accuracy

A new study from the Journal of Artificial Intelligence Research (JAIR) introduces Input‑Skip Latent Binary Bayesian Neural Networks (ISLaB), a variant of latent binary Bayesian neural networks (LBBNNs) that can prune connections during training and allow inputs to skip layers or be omitted entirely. The authors—Eirik Høyheim, Lars Skaaret‑Lund, Solve Sæbø, and Aliaksandr Hubin—published their findings on July 1, 2026 in JAIR’s latest issue. Their approach achieves over 99 % reduction in network density for small models and more than 99.9 % for larger ones while still delivering accurate predictions and uncertainty estimates ¹.

The core innovation lies in letting covariates decide whether they should propagate forward or bypass subsequent layers. This dynamic routing simplifies the architecture and shows which features matter most at each layer. By learning which inputs are most relevant at every layer, ISLaB can uncover causal relationships between features that standard dense networks obscure. On the MNIST handwritten digit dataset the method reaches 97 % classification accuracy with just 935 weights, matching state-of-the-art calibration levels despite using only about one thousand parameters total—a dramatic drop compared to conventional deep learning architectures used today ¹.

Expert analysis highlights that this breakthrough builds on earlier work by MacKay and Tipping from 1998, who first introduced Bayesian networks as a way of quantifying uncertainty. Earlier Bayesian neural networks struggled with computational complexity because of high parameter counts; ISLaB addresses this challenge by drastically reducing model size without sacrificing predictive performance. The ability to selectively prune connections during training also improves interpretability, allowing researchers to see directly which inputs matter where in the network.

Beyond academic interest, the technique offers practical benefits. A model with fewer than a thousand weights requires far less memory, lower energy consumption, and faster inference times—critical for deploying AI on edge devices or large‑scale cloud services where cost savings could reach millions per year according to internal estimates within Webb Local AI lab. Maintaining accurate uncertainty estimates means the system can flag low‑confidence predictions, improving safety in applications such as medical diagnosis or autonomous driving.

The study suggests future research directions: extending input‑skip LBBNNs to other datasets beyond MNIST, exploring how the approach scales with deeper architectures, and integrating it into existing probabilistic programming frameworks. As artificial intelligence systems continue to grow in size and complexity, methods that keep models lean while preserving accuracy will be increasingly valuable.