Valeriia Cherepanova
I am a postdoctoral scientist at Amazon AWS Responsible AI. Currently, I am working on problems related to safety and security of large language models. The development of LLMs introduces great capabilities, but we continue to identify new vulnerabilities and failure modes of these systems, and we still do not understand their inner workings. In my research I explore robustness of language models to out-of-domain inputs, and jailbreak attacks, detectablity of AI generated text, and biases in these models. Another area of interest to me is deep learning for tabular problems, my goal is to advance machine learning in tabular domain beyond traditional decision tree models.
Prior to joining Amazon I did my PhD at the University of Maryland, where I was advised by Prof. Tom Goldstein and worked on various topics in machine learning ranging from fairness and robustness of deep neural networks to tabular deep learning. During my PhD I did two summer internships at Amazon, in 2021 I developed NLP solution to monitor integrity and transparency of third-party Alexa Skills at Alexa Monitoring team and in 2022 I worked on improving Alexa Voice Search on FireTV at Alexa Entertainment team.
I am always looking for new collaborations, feel free to text me on LinkedIn or send me an email!
news
Jan 24, 2024 | Spotting LLMs With Binoculars is now on arxiv, check our demo on huggingface! |
---|---|
Sep 21, 2023 | A Performance-Driven Benchmark for Feature Selection got accepted to NeurIPS 23 |
Jul 27, 2023 | I defended my PhD! |
selected publications
- Lowkey: Leveraging adversarial attacks to protect social media users from facial recognitionInternational Conference on Learning Representations, 2021
- Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated TextarXiv preprint arXiv:2401.12070, 2024
- Transfer learning with deep tabular modelsInternational Conference on Learning Representations, 2023
- A Performance-Driven Benchmark for Feature Selection in Tabular Deep LearningAdvances in Neural Information Processing Systems, 2023
- Unraveling meta-learning: Understanding feature representations for few-shot tasksIn International Conference on Machine Learning, 2020
- TuneTables: Context Optimization for Scalable Prior-Data Fitted NetworksarXiv preprint arXiv:2402.11137, 2024