Scholar
Marius-Constantin Dinu
Google Scholar ID: vh39WvYAAAAJ
PhD, Institute of Machine Learning, JKU, Linz, ExtensityAI Research Scientist
Citations & Impact
Citations
284
H-index
8
i10-index
8
Publications
19
Co-authors
0
Publications
Resume
Academic Achievements
- 1. Published work on integrating symbolic methods with large language models, introducing a new API that bridges classical programming and differentiable programming.
- 2. Proposed a novel approach to address parameter choice issues in unsupervised domain adaptation, which outperformed existing benchmark methods across several types of datasets.
- 3. Introduced the concept of reactive exploration in the context of lifelong reinforcement learning, demonstrating through experiments that policy gradient methods perform better under such conditions.
Research Experience
- 1. Worked on integrating symbolic methods and large language models, proposing a new framework that uses neural networks, specifically LLMs, at its core, and composes operations based on task-specific zero-shot or few-shot prompting.
- 2. Researched the problem of choosing algorithm hyper-parameters in unsupervised domain adaptation, proposing an extension of weighted least squares to vector-valued functions, and conducted a large-scale empirical comparative study across multiple datasets.
- 3. Explored reactive exploration strategies for dealing with continuous domain shifts in lifelong reinforcement learning, showing that policy gradient methods adapt more quickly to distribution shifts than Q-learning.
Background
- Research interests include a neuro-symbolic perspective on Large Language Models (LLMs), addressing parameter choice issues in unsupervised domain adaptation, and coping with non-stationarity in lifelong reinforcement learning. The main focus is on combining symbolic approaches with deep learning techniques to solve complex problems.