NeuralBayesianNetworks: A PyTorch-Native Bayesian Network Library Enabling Full GPU Acceleration for Probabilistic Inference

A fully PyTorch-based Bayesian network library that models each conditional probability distribution as a learnable neural network module, supporting batch tensor operations and end-to-end GPU acceleration, with inference speed 9-22 times faster than traditional libraries.

• Original Author/Maintainer: Giovanni Briglia
• Source Platform: GitHub
• Original Title: NeuralBayesianNetworks
• Original Link: https://github.com/Giovannibriglia/NeuralBayesianNetworks
• Release Date: June 3, 2026

---

As a core representative of probabilistic graphical models, Bayesian networks have been applied in fields such as uncertainty reasoning, causal analysis, and decision support for decades. However, traditional implementations like pgmpy are often limited by Python loops and CPU computation when dealing with large-scale data, making it difficult to meet the speed and scale requirements of modern machine learning.

NeuralBayesianNetworks (NBN for short) emerged as a solution—it is a fully PyTorch-based Bayesian network library that models the Conditional Probability Distribution (CPD) of each node as a learnable neural network module, enabling true end-to-end GPU acceleration. Developer Giovanni Briglia positions it as "the GPyTorch of Bayesian networks", meaning it brings native tensor operations and automatic differentiation capabilities to Bayesian networks, just as GPyTorch revolutionized Gaussian processes.

---

NBN's design philosophy is completely different from traditional Bayesian network libraries. In traditional implementations, conditional probability tables are static data structures, and the inference process involves a lot of discrete lookups and Python-level loops. In contrast, NBN treats each conditional probability distribution as a learnable neural network module.

Specifically, each node in NBN carries a learnable, batch-processing-supported, GPU-resident conditional distribution. This means:

• Conditional Probability Distribution as Neural Network: Whether it's a classification table for discrete variables, a mixture density network for continuous variables, or a complex Gaussian process, all can serve as the conditional probability mechanism for nodes
• Query as Tensor Operation: All inference queries are expressed as batch tensor operations, natively supporting PyTorch's automatic differentiation and GPU acceleration
• End-to-End Learnable: Both parameter learning and structure learning can be done via gradient descent, without the need for traditional MLE or EM algorithms

This architecture allows NBN to seamlessly integrate into modern deep learning workflows, enabling researchers to use the familiar PyTorch API to build, train, and infer complex probabilistic graphical models.

---

NBN shows significant advantages in performance. According to the author's benchmark tests on an RTX 4070 laptop (version v0.6c-d), NBN achieves amazing acceleration compared to traditional libraries:

On continuous linear Gaussian networks, NBN is 9-22 times faster than pgmpy:

• When the number of nodes n=10, NBN takes only 1.9 ms, while pgmpy takes 42 ms (22x acceleration)
• When n=1000, NBN takes 0.72 seconds, while pgmpy takes 8.5 seconds (12x acceleration)
• More importantly, NBN maintains accuracy while improving speed—the difference in Wasserstein-1 distance from pgmpy remains within 0.02 at all times

In the parameter learning task for discrete Bayesian networks, NBN shows stronger scalability:

• When the number of nodes ≥50, NBN's Total Variation (TV) distance stabilizes at around 0.14
• In contrast, pgmpy's MLE method saturates at TV≈0.34
• The quality gap starts at 0.10 vs 0.25 when n=50, and continues to 0.146 vs 0.340 when n=1000

This advantage comes from NBN's gradient optimization method, which can break through the sample complexity limitations of traditional MLE methods and maintain learning quality on larger-scale problems.

NBN is currently the only library that can natively handle large-scale hybrid (continuous-discrete) networks. In benchmark tests, only Pyro can support hybrid network inference via importance sampling, while pgmpy, GPyTorch, and pomegranate do not have applicable hybrid baselines. NBN-hybrid can effectively handle hybrid networks across all tested scales (10 to 1000 nodes).

---