About Me
Hi! I’m Victoria Zhanqi Zhang (张展旗), a Ph.D. candidate in Computer Science at UC San Diego, supported by the HDSI Ph.D. Fellowship. I’m co-advised by Dr. Mikio Aoi and Dr. Gal Mishne. I earned dual B.S. in Computer Science and Electrical Engineering and a M.S. in Computer Science from Washington University in St. Louis, where I worked with Dr. Carlos Ponce as an undergraduate researcher.
I develop generative and self-supervised AI models that reveal structure and dynamics in complex, large-scale systems such as the brain and behavior. See my research.
I enjoy traveling, art, and animals, and share my home with birds. See my art portfolio.
For more details, see CV.
News
- Hire Me — Open to full-time Research Scientist roles in 2026. Connect with me on LinkedIn!
- New preprint The Position Curse: LLMs Struggle to Locate the Last Few Items in a List, now available on arXiv.
- Latest work on adversarial domain adaptation ALIGN, accepcted at UAI.
- My work, BEHAVE: Behavioral Ethology for Human Assessment via Variational Encoding, has been accepted at the NeurIPS 2025 Workshop on Data on Brain and Mind (DBM).
- My paper, Brain Feature Maps Reveal Progressive Animal-Feature Representations, is published in Science Advances and featured as “When Neurons Discover on Their Own” by Harvard Brain Institute News.
- Check out my preprint on Behavioral Dynamics in Bipolar Disorder on medRxiv.
- Co-authored paper Arousal as a universal embedding for spatiotemporal brain dynamics now published in Nature.
Research
Machine Learning I develop robust machine learning methods for large-scale spatiotemporal data, studying how self-supervised and generative models learn hierarchical, multi-scale structure. My work includes developing machine learning methods for robust generalization under distribution shift, enabling effective few-shot and zero-shot adaptation for online sequence-to-text decoding (paper), developing multimodal foundation models + language models, with a focus on scalable training, efficient on-device optimization, and robust real-time decoding under real-world kinematic variation for the Meta Neural Band.
Computational Neuroscience I study how to neural representations and learning dynamics encode across brain and manifest in behaviors. I showed how hierarchical feature representations arise along the primate ventral visual stream (paper). I developed latent representation + temporal modeling methods for long-form human video understanding, leveraging unsupervised learning and state-space modeling to uncover subtle behavioral structure from noisy, weakly supervised video data in bipolar disorder. (preprint).
Industry Experience In 2024, I developed CLIP-style multimodal foundation models for neural interfaces, aligning EMG, IMU, and vision into a shared embedding space to improve cross-device generalization and downstream performance across multiple tasks, such as pose estimation, at Meta. In 2025, I focused on production-scale model training and on-device optimization for handwriting recognition, balancing latency, accuracy, and robustness for real-world deployment. This work contributed to Meta Neural Band production, which captures electrical activity from subtle forearm muscle movements for intuitive, gesture-based input, and live demoed at Meta Connect 2025. The broader product effort was recognized as one of TIME’s Best Inventions of 2025.