Hanna Tseran
Hanna Tseran

Postdoctoral Researcher

RIKEN AIP | Tokyo, Japan

Hi! I’m Hanna. I investigate the fundamentals of contemporary neural networks to make AI systems more capable and understandable.

Research Interests
  • Theoretical foundations
    Deep learning theory, mathematical approaches for model development and understanding
  • Modern neural architectures
    Design and study of contemporary models, such as LLMs and other transformer-based networks
  • Model behavior analysis
    Investigating phenomena like emergent abilities and in-context learning
  • Model performance
    Optimization analysis, efficiency improvements
About Me

I’m a postdoctoral researcher at RIKEN AIP in Tokyo, Japan. Our lab is a part of RIKEN and the University of Tokyo, so I’m involved in both.

I focus on deep learning theory, and I’m passionate about understanding the fundamental principles behind modern neural networks. I believe this is the best way to build safe and capable AI systems in the long run. To this end, I draw on methods from various branches of mathematics in my work. Currently, my research centers on the theory behind transformer-based models, such as LLMs.

I’m originally from Belarus but moved to Japan a while ago. I’ve also lived in Germany and the UK. I have many hobbies, including yoga and hiking, and more creative ones like painting, embroidery, and cooking. I speak English, Japanese, Belarusian, and Russian.

I’m always happy to have an interesting conversation, so please feel free to contact me at hanna.tseran@gmail.com.

News

📅 WISJ Machine Learning Summer School volunteering

This summer, I’m volunteering as a mentor and a teaching assistant at the Machine Learning Summer School for Scientists 🔗, which started today and is organized by the awesome Women in Science Japan community.

🎉 This website is up

Welcome! I plan to publish here details of my work and news.

KAKENHI 2025

I was awarded a JSPS Grant-in-Aid for Early-Career Scientists (KAKENHI) to support my work on a theoretical framework for in-context learning development.

Publications
(2024). Mildly Overparameterized ReLU Networks Have a Favorable Loss Landscape. In TMLR.
(2023). Expected Gradients of Maxout Networks and Consequences to Parameter Initialization. In ICML.
(2021). On the Expected Complexity of Maxout Networks. In NeurIPS.
(2018). Natural Variational Continual Learning. In NeurIPS Workshop on Continual Learning.
(2017). Memory augmented neural network with Gaussian embeddings for one-shot learning. In NeurIPS Workshop on Bayesian Deep Learning.