If you’re working on probabilistic models, decision making under uncertainty, or neurosymbolic methods, I'd love for you to check it out, try it, and send feedback!⁣

This package integrates methods from multiple publications (AAAI, NeurIPS, AISTATS, and more), all unified into a single, easy-to-use codebase.⁣
⁣
Whether you're an ML researcher or systems builder—𝐍𝐞𝐮𝐏𝐈 gives you tools to accelerate and scale inference in these models.⁣

🔧 What’s Inside?⁣
⁣
* 🧩 Modular design with plug-and-play support for PGMs and neural solvers⁣
* 🔁 ITSELF: our test-time refinement method
⁣* ⚡ Fast, extensible, and backed by a Cython-powered backend⁣
* 📦 Built for both probabilistic graphical models and probabilistic circuits

NeuPI is a PyTorch-based framework for neural probabilistic inference. It introduces a self-supervised training paradigm where the probabilistic model itself supervises the neural network, eliminating the need for annotated data.⁣

⁣

I'm thrilled to release 𝐍𝐞𝐮𝐏𝐈, a Python library that answers this question with a resounding 𝐲𝐞𝐬.⁣

🔗 𝐃𝐨𝐜𝐬: neupi.readthedocs.io/en/latest/
💻 𝐆𝐢𝐭𝐇𝐮𝐛: github.com/Shivvrat/NeuPI

#AI #MachineLearning #ProbabilisticModels
#AISTATS2025 #NeuralNetworks

I am grateful for the opportunity to work with such talented collaborators and mentors! A special thank you to 𝗣𝗿𝗼𝗳. Vibhav Gogate for his constant support and mentorship.

Looking forward to presenting our work at [AISTATS] International Conference on Artificial Intelligence and Statistics 𝟮𝟬𝟮𝟱! 😊

📊 𝗞𝗲𝘆 𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆𝘀:
Our approach significantly improves accuracy and scalability across various applications, making it more practical and impactful for real-world problems.

👥 𝗖𝗼-𝗮𝘂𝘁𝗵𝗼𝗿𝘀: Dr. Tahrima Rahman and Prof. Vibhav Gogate

2. 𝗕𝗲𝘁𝘁𝗲𝗿 𝗽𝗿𝗲𝗱𝗶𝗰𝘁𝗶𝗼𝗻: We develop two techniques to improve prediction accuracy through more effective discretization:
- A method that uses an "oracle" to resolve uncertain variables by using other strong predictions.
- A scoring-based method to find the best nearby discrete solution.

🚀 𝗢𝘂𝗿 𝗔𝗽𝗽𝗿𝗼𝗮𝗰𝗵:
We introduce a novel solution that solves these challenges:
1. 𝗘𝗻𝗵𝗮𝗻𝗰𝗲𝗱 𝗳𝗲𝗮𝘁𝘂𝗿𝗲 𝗲𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝘀: We incorporate the structure and parameters of the PGM, making the neural network smarter and more effective.

In recent years, 𝗻𝗲𝘂𝗿𝗮𝗹 𝗻𝗲𝘁𝘄𝗼𝗿𝗸𝘀 have been used to generate these predictions, but existing methods face two key challenges:
1. 𝗟𝗶𝗺𝗶𝘁𝗲𝗱 𝘂𝘀𝗲 𝗼𝗳 𝗺𝗼𝗱𝗲𝗹 𝗸𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲
2. 𝗜𝗺𝗽𝗿𝗲𝗰𝗶𝘀𝗲 𝗽𝗿𝗲𝗱𝗶𝗰𝘁𝗶𝗼𝗻𝘀

Solving it efficiently for large and complex systems has been a huge challenge—until now!

🤔 𝗪𝗵𝗮𝘁 𝗶𝘀 𝘁𝗵𝗶𝘀 𝗿𝗲𝘀𝗲𝗮𝗿𝗰𝗵 𝗮𝗯𝗼𝘂𝘁?
Imagine you are given incomplete data and need to predict the most likely scenario that explains it. For example, in healthcare, given symptoms (evidence), doctors may want to infer the most probable diagnosis. This type of problem is called the 𝗠𝗣𝗘 query.

I’m looking forward to engaging discussions, insightful questions, and connecting with fellow researchers. If you’re attending NeurIPS, stop by my sessions—I’d love to chat! 🚀

#NeurIPS2024 #MachineLearning #ComputerVision #ProbabilisticModels #ErrorRecognition #AIResearch #MLResearch

🥗 𝐏𝐚𝐩𝐞𝐫 𝟐: 𝐶𝑎𝑝𝑡𝑎𝑖𝑛𝐶𝑜𝑜𝑘4𝐷: 𝐴 𝐷𝑎𝑡𝑎𝑠𝑒𝑡 𝑓𝑜𝑟 𝑈𝑛𝑑𝑒𝑟𝑠𝑡𝑎𝑛𝑑𝑖𝑛𝑔 𝐸𝑟𝑟𝑜𝑟𝑠 𝑖𝑛 𝑃𝑟𝑜𝑐𝑒𝑑𝑢𝑟𝑎𝑙 𝐴𝑐𝑡𝑖𝑣𝑖𝑡𝑖𝑒𝑠
📅 𝐏𝐨𝐬𝐭𝐞𝐫 𝐒𝐞𝐬𝐬𝐢𝐨𝐧: Friday, Dec 13, 2024, 11:00 AM - 2:00 PM
📍 𝐋𝐨𝐜𝐚𝐭𝐢𝐨𝐧: West Ballroom A-D (#5308)

🧠 𝐏𝐚𝐩𝐞𝐫 𝟏 (𝐒𝐩𝐨𝐭𝐥𝐢𝐠𝐡𝐭): 𝐴 𝑁𝑒𝑢𝑟𝑎𝑙 𝑁𝑒𝑡𝑤𝑜𝑟𝑘 𝐴𝑝𝑝𝑟𝑜𝑎𝑐ℎ 𝑓𝑜𝑟 𝐸𝑓𝑓𝑖𝑐𝑖𝑒𝑛𝑡𝑙𝑦 𝐴𝑛𝑠𝑤𝑒𝑟𝑖𝑛𝑔 𝑀𝑜𝑠𝑡 𝑃𝑟𝑜𝑏𝑎𝑏𝑙𝑒 𝐸𝑥𝑝𝑙𝑎𝑛𝑎𝑡𝑖𝑜𝑛 𝑄𝑢𝑒𝑟𝑖𝑒𝑠 𝑖𝑛 𝑃𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑠𝑡𝑖𝑐 𝑀𝑜𝑑𝑒𝑙𝑠
📅 𝐏𝐨𝐬𝐭𝐞𝐫 𝐒𝐞𝐬𝐬𝐢𝐨𝐧: Thursday, Dec 12, 2024, 11:00 AM - 2:00 PM
📍 𝐋𝐨𝐜𝐚𝐭𝐢𝐨𝐧: East Exhibit Hall A-C (#4104)