David Debot

Just under 10 days left to submit your latest endeavours in #tractable probabilistic models!

Join us at TPM @auai.org #UAI2025 and show how to build #neurosymbolic / #probabilistic AI that is both fast and trustworthy!

We developed a library to make logical reasoning embarrasingly parallel on the GPU.

For those at ICLR 🇸🇬: you can get the juicy details tomorrow (poster #414 at 15:00). Hope to see you there!

If you're at #AAAI2025, come check out our demo on neurosymbolic reinforcement learning with probabilistic logic shields 🤖 Tomorrow (Sat, March 1) from 12:30–2:30 PM during the poster session 💻

We all know backpropagation can calculate gradients, but it can do much more than that!

Come to my #AAAI2025 oral tomorrow (11:45, Room 119B) to learn more.

🔥 Can AI reason over time while following logical rules in relational domains? We will present Relational Neurosymbolic Markov Models (NeSy-MMs) next week at #AAAI2025! 🎉

📜 Paper: arxiv.org/pdf/2412.13023
💻 Code: github.com/ML-KULeuven/...

🧵⬇️

See you at #AAAI2025!

Site: dtai.cs.kuleuven.be/projects/nes...

Video: youtu.be/3uLVxwlcSQc?...

@daviddebot.bsky.social, @gabventurato.bsky.social, @giuseppemarra.bsky.social, @lucderaedt.bsky.social

#ReinforcementLearning #AI #MachineLearning #NeurosymbolicAI
(8/8)

Open-source & easy to use!
🔷 Code: github.com/ML-KULeuven/...
🔷 Based on MiniHack & Stable Baselines3
🔷 Define new shields in just a few lines of code!

🚀 Let’s make RL safer & smarter, together!
(7/8)

Want to try it yourself? 🎮

Use our interactive web demo!
🔷 Modify environments (add lava, monsters!)
🔷 Test shielded vs. non-shielded agents

🖥️ Play with it here: dtai.cs.kuleuven.be/projects/nes...
(6/8)

Why does this matter?
🔷 Faster training ⌛
🔷 Safer exploration 🔒
🔷 Better generalization 🌍
(5/8)

How does it work? 🤔🛡️

The shield:
✅ Exploits symbolic data from sensors 🌍
✅ Uses logical rules 📜
✅ Prevents unsafe actions 🚫
✅ Still allows flexible learning 🤖

A perfect blend of symbolic reasoning & deep learning!
(4/8)

Enter MiniHack, our demo's testing ground! 🏰🗡️

There, RL agents face:
✅ Lava cliffs & slippery floors
✅ Chasing monsters
✅ Locked doors needing keys

Findings:
🔷 Standard RL struggles to find an optimal, safe policy.
🔷 Shielded RL agents stay safe & learn faster!
(3/8)

Deep RL is powerful, but...
⚠️ It can take dangerous actions
⚠️ It lacks safety guarantees
⚠️ It struggles with real-world constraints

Yang et al.'s probabilistic logic shields fix this, enforcing safety without breaking learning efficiency! 🚀
(2/8)

🚀 Do you care about safe AI? Do you want RL agents that are both smart & trustworthy?

At #AAAI2025, we present our demo for neurosymbolic RL—combining deep learning with probabilistic logic shields for safer, interpretable AI in complex environments. 🏰🔥
🧵👇
(1/8)

YouTube video by David Debot Interpretable Concept-Based Memory Reasoning - NeurIPS 2024

A short overview video can be found on YouTube: youtu.be/CgSDhQKESD0?...

#NeurIPS2024

Or check out our Medium post: 👉 medium.com/@pyc.devteam... (7/7)

With CMR, we’re reaching the sweet spot of accuracy and interpretability. Check it out at our poster at #NeurIPS2024! 👉 neurips.cc/virtual/2024... (6/7)

During training, CMR learns embeddings as latent representations of logic rules, and a neural rule selector identifies the most relevant rule for each instance. Due to a clever factorization and rule selector, inference is linear in the number of concepts and rules. (5/7)

CMR makes a prediction in 3 steps:
1) Predict concepts from the input
2) Neurally select a rule from a memory of learned logic rules ➨ Accuracy
3) Evaluate the selected rule with the concepts to make a final prediction ➨ Interpretability (4/7)

CMR has:
⚡ State-of-the-art accuracy that rivals black-box models
🚀 Pure probabilistic semantics with linear-time exact inference
👁️ Transparent decision-making so human users can interpret model behavior
🛡️ Pre-deployment verifiability of model properties (3/7)

CMR is our latest neurosymbolic concept-based model. A proven 𝘶𝘯𝘪𝘷𝘦𝘳𝘴𝘢𝘭 𝘣𝘪𝘯𝘢𝘳𝘺 𝘤𝘭𝘢𝘴𝘴𝘪𝘧𝘪𝘦𝘳 irrespective of the concept set, CMR achieves near-black-box accuracy by combining 𝗿𝘂𝗹𝗲 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴 and 𝗻𝗲𝘂𝗿𝗮𝗹 𝗿𝘂𝗹𝗲 𝘀𝗲𝗹𝗲𝗰𝘁𝗶𝗼𝗻! (2/7)

🚨 Interpretable AI often means sacrificing accuracy—but what if we could have both? Most interpretable AI models, like Concept Bottleneck Models, force us to trade accuracy for interpretability.

But not anymore, due to Concept-Based Memory Reasoner (CMR)! #NeurIPS2024 (1/7)