Projects

Implementation and evaluation of structured Multivariate Autoregressive (MVAR) and Latent Dynamical System (LDS) models with low-rank inputs for analyzing neural time series data. Developed tailored parameter estimation algorithms and a nested cross-validation framework to robustly recover latent dynamics and structured variability across trials. This work forms the foundation for my ongoing PhD research.

Integration of two neural dynamics modeling frameworks, Low Tensor Rank RNNs and iLQR-VAE to analyze learning-induced changes in neural connectivity by inferring inputs and connectivity changes in nonlinear dynamical systems from neural data. Work involved wrestling with fundamental identifiability challenges in separating intrinsic dynamics from external forcing in non-linear systems. This work forms the foundation for my ongoing PhD research.

Multi-label safety classification system for imbalanced datasets, integrating active learning with human annotator feedback. Fine-tuned transformer models (T5, LLaMA, BERT) using LoRA and 8-bit quantization, comparing generative multi-label (seq2seq) and discriminative binary classification approaches. Exploration of LLM-assisted annotation pipelines (LLMaAA methodology) combining prompt engineering with active learning strategies for scalable data collection.

Machine learning pipeline for identifying annotation quality issues in safety-critical text classification datasets with limited labeled data. Implemented confidence learning-based system to detect mislabeled annotations with systematic false positive/false negative analysis. Evaluation framework combining label confidence scores, ROC-AUC, and precision/recall metrics. Explored multiple text embedding approaches (Cohere, Sentence Transformers) and clustering methods for annotation pattern analysis.