Input‑skip LBBNNs cut parameters >99%

Eirik Høyheim, Lars Skaaret-Lund, Solve Sæbø, and Aliaksandr Hubin introduced a new class of latent binary Bayesian neural networks called input‑skip LBBNNs that allow covariates to skip any succeeding layer or be excluded entirely, simplifying network structure and clarifying how inputs influence predictions. This approach was published on 2026/07/01 in the Journal of Artificial Intelligence Research.¹

The authors show that input‑skip LBBNNs achieve dramatic reductions in parameter count compared with standard LBBNNs. For small networks they reduce density by over 99 %, and for larger ones by over 99.9 %. Despite these reductions, the models retain high predictive accuracy and uncertainty quantification. On the MNIST digit recognition task an input‑skip LBBNN reached 97 % classification accuracy using just 935 weights, achieving great calibration.¹

They emphasize that the use of latent binary variables introduces structural sparsity that can be interpreted similarly to feature selection techniques from earlier decision tree work. This explainability is especially valuable in high‑stakes domains such as healthcare, where understanding which connections drive decisions matters. However, they note that calibration quality remains an open question: poorly calibrated uncertainty estimates could lead to overly confident but incorrect predictions even when point estimates are accurate.¹

These findings suggest a significant step toward efficient inference at scale, as modest improvements translate into substantial cost savings given current large language model sizes. The approach also highlights the importance of balancing interpretability, efficiency, and reliable uncertainty estimation in next‑generation AI systems.