Pavel Czempin

Pavel Czempin

CS PhD Student

University of Southern California

Biography

Hi There! I’m a PhD student in Computer Science at the University of Southern California. I work at the USC Learning and Interactive Robot Autonomy Lab, advised by Erdem Biyik. I have previously done research at the Center for Human-Compatible AI at UC Berkeley and Fraunhofer AISEC and have a Master’s degree from the Technical University of Munich.

Feel free to reach out, I’m interested in collaborating!

Interests
  • Reinforcement Learning from Human Feedback
  • Generalization in AI
  • Robotics
  • Trustworthy AI
Education
  • PhD in Computer Science (Artificial Intelligence)

    University of Southern California

  • MSc in Computer Science (Artificial Intelligence), 2023

    Technical University of Munich

  • BSc in Engineering Science, 2020

    Technical University of Munich

Research Experience

 
 
 
 
 
Center for Human-Compatible AI
Research Intern
May 2024 – Aug 2024 UC Berkeley
  • Mentor: Rachel Freedman
 
 
 
 
 
Center for Human-Compatible AI
Visiting Student Researcher
Jan 2022 – Aug 2022 UC Berkeley
  • Continued research on defenses against adversarial reinforcement learning policies.
  • Research on interpretability for learned reward models.
  • In Collaboration with Adam Gleave, Daniel Filan
 
 
 
 
 
Center for Human-Compatible AI
Research Collaborator
May 2021 – Dec 2021 UC Berkeley (remote)
  • Research on defenses against adversarial reinforcement learning policies.
  • Supervised by Adam Gleave
 
 
 
 
 
Research Assistant
May 2020 – Aug 2021 Munich, Germany
  • Implemented and compared audio deepfake detection architectures
  • Supervised by Nicolas Müller

Publications

(2024). In-Context Generalization to New Tasks From Unlabeled Observation Data. In 1st ICL Workshop at ICML 2024.

PDF

(2022). Does Audio Deepfake Detection Generalize?. In Proc. Interspeech 2022.

PDF Cite DOI

(2022). Reducing Exploitability with Population Based Training. In AdvML Frontiers at ICML.

PDF Cite Code

(2021). Speech is Silver, Silence is Golden: What do ASVspoof-trained Models Really Learn?. In ASVspoof2021.

PDF Cite DOI