Samuel Teuber

Doctoral Researcher @KITinformatik (@KITKarlsruhe) | Formal Methods for Software & AI (Focus on CPS and Fairness Verification)

[Samuel Teuber @teuber_dev]

You want to ensure that your neural network *never* crashes your control system?

Our (now accepted 🥳) #NeurIPS paper introduces:
- Reusing control theory for NN verification
- Verifying *nonlinear arithmetic* specs on NNs

This guarantees your NN won't behave like this (1/12):

[Samuel Teuber @teuber_dev]

I decided I'll try out the other option...

[Ilya Sergey @ilyasergey]

That was surprisingly accurate.

[Samuel Teuber @teuber_dev]

If anyone in Edinburgh is interested in learning more about this work:

I will give a talk about it at the LFCS Seminar @InfAtEd on Tuesday, 29th of October, at 4 pm.

https://t.co/eDMGWnQRzG

[Samuel Teuber @teuber_dev]

You want to ensure that your neural network *never* crashes your control system?

Our (now accepted 🥳) #NeurIPS paper introduces:
- Reusing control theory for NN verification
- Verifying *nonlinear arithmetic* specs on NNs

This guarantees your NN won't behave like this (1/12):

[Samuel Teuber @teuber_dev]

@AndrePlatzer @SCSatCMU @KITinformatik Our verification tool for polynomial arithmetic specifications (possibly with prepositionally complicated structure) for (ReLU) neural networks is also open source and can be found here:
https://t.co/iv21AYPgHN

[Samuel Teuber @teuber_dev]

You want to ensure that your neural network *never* crashes your control system?

Our (now accepted 🥳) #NeurIPS paper introduces:
- Reusing control theory for NN verification
- Verifying *nonlinear arithmetic* specs on NNs

This guarantees your NN won't behave like this (1/12):

[Samuel Teuber @teuber_dev]

This project was joint work with Stefan Mitsch and @AndrePlatzer. We started working on this project during my visit to @SCSatCMU and continued working on it @KITinformatik. I'm looking forward to presenting it at #neurips2024 in Vancouver.
Preprint: https://t.co/ryA0jumu8X

[Brendan Dolan-Gavitt @moyix]

@alpha_convert Bet you’re feeling pretty vindicated today huh

[Samuel Teuber @teuber_dev]

@moyix Sometimes it even tells you what’s wrong:
Recently, when I ran an experiment with GPT 4o, I noticed it responded something like „I’m afraid I can’t see the prior conversation, but let me take an educated guess“ – indeed, it was not seeing most of the chat history due to a bug…

[KASTEL @KASTEL_SRL]

Please find the registration form here: https://t.co/3fDVmalAFM

[Samuel Teuber @teuber_dev]

Correction: It's actually poster number 205

[Samuel Teuber @teuber_dev]

Want to learn how we can harness well-established program analysis techniques to analyse fairness properties?
Drop by my #AAAI poster (number 204) this evening!
#AAAI2024 #AAAI24

[Samuel Teuber @teuber_dev]

Our paper "An Information-Flow Perspective on Algorithmic Fairness" was accepted for AAAI 2024:

We looked at similarities between Algorithmic Fairness and properties checked by Information Flow Analysis Tools.
Outcome: We can verify a software's fairness via Information Flow
🧵