Samuele Bortolotti

In collaboration with @ema-ridopoco.bsky.social Tommaso Carraro @paolomorettin.bsky.social @emilevankrieken.com @nolovedeeplearning.bsky.social @looselycorrect.bsky.social @andreapasserini.bsky.social

Not All Neuro-Symbolic Concepts Are Created Equal: Analysis and Mitigation of Reasoning Shortcuts Neuro-Symbolic (NeSy) predictive models hold the promise of improved compliance with given constraints, systematic generalization, and interpretability, as they allow to infer labels that are consiste...

Want to know more?

1️⃣ Learn more about RSs: Why they appear, their root causes, and mitigation: arxiv.org/abs/2305.19951

2️⃣ Make NeSy models aware of their shortcuts: arxiv.org/abs/2402.12240

For other details regarding rsbench, datasets, and experiments, check the links below:

Website: unitn-sml.github.io/rsbench/
Paper: openreview.net/forum?id=5Vt...
GitHub: github.com/unitn-sml/rs...

Easy to set up and use!

1️⃣ Configurable: can be easily configured with YAML/JSON files.
2️⃣ Intuitive: straightforward to use:

📊 8 challenging tasks, all with predefined settings.

3 new benchmarks:
🔢 MNMath for arithmetic reasoning
🛃 MNLogic for SAT-like problems
🚖 SDD-OIA, a synthetic self-driving task!

They can all be made easier or harder with our data generator!

🧪 Test your models!

- 🌍 Evaluate concepts in in- and out-of-distribution scenarios.
- 🎯 Ground-truth concept annotations are available for all tasks.
- 📊 Visualize how your models handle different learning & reasoning tasks!

🔍 rsbench allows you to:

- 🧮 Run algorithmic, logical, and high-stakes tasks w/ known reasoning shortcuts (RSs).
- 📊 Eval concept quality via F1, accuracy & concept collapse.
- 🛠️ Easily customize the tasks and count RSs a priori using our countrss tool!

🤔 What are reasoning shortcuts?

NeSy models might learn wrong concepts but still make perfect predictions!

Example: A self-driving car 🚗 stops in front of a 🚦🔴 or a 🚶. Even if it confuses the two, it outputs the right prediction!

🌐 rsbench allows you to evaluate the concepts learned by:

1️⃣ Neuro-Symbolic models (#NeSy)
2️⃣ Concept Bottleneck Models (#CBMs)
3️⃣ Black-box Neural Networks (NNs*)
4️⃣ Vision-Language Models (#VLMs*)

* through post-hoc concept-based explanations (e.g., TCAV)

📣 Does your model learn high-quality #concepts, or does it learn a #shortcut?

Test it with our #NeurIPS2024 dataset & benchmark track paper!

rsbench: A Neuro-Symbolic Benchmark Suite for Concept Quality and Reasoning Shortcuts

What's the deal with rsbench? 🧵

by @ema-ridopoco.bsky.social @looselycorrect.bsky.social @andreapasserini.bsky.social @samubortolotti.bsky.social

eg

👉 proceedings.neurips.cc/paper_files/...

👉 openreview.net/forum?id=pDc...

👉 unitn-sml.github.io/rsbench/