Research
CBIRD
A complex-valued BERT language model for hybrid classical/quantum computing
-
Undergraduate Research Assistant, AI & NLP Lab
Quantum AI Study Group; Advisor: Prof. Damir Cavar, Indiana University Bloomington
Jan 2026 – Present
- Co-developing CBIRD, a complex-valued BERT language model designed for hybrid classical/quantum computing, alongside a real-valued twin model for controlled comparison.
- Adapted complex-valued whitening normalization from Deep Complex Networks (Trabelsi et al.) for use as layer normalization in a transformer architecture.
- Diagnosed a failure mode in which Born-rule output probabilities collapsed toward the unigram distribution; introduced a tunable per-layer escort (β) parameter that enabled the model to train.
- Design and run experiments comparing complex- and real-valued models across activation functions, normalization methods, and parameter counts, evaluated on embedding quality and downstream classification and sentiment analysis.
