Scalable Neural Models for Your Needs

I specialize in developing accurate and interpretable neural networks, challenging the common misconception that model accuracy and interpretability are inversely correlated. I demonstrate that models with proper representation can be both highly accurate and fully interpretable. My work provides interpretation and verification for domain experts, using domain-guided losses and orthogonality to address tasks in machine learning or natural language tasks, and computational sciences.

My PhD thesis focuses on developing accurate and fair neural models for Natural Language tasks, particularly toxicity detection. It addresses the challenge of varying toxic language across demographic groups by promoting group-specific diversity measures. I tackle the risks of overfitting to majority groups and the fairness concerns that arise with "one-size-fits-all" models. My work uses multi-task learning for group-targeted toxicity detection and fairness loss functions to ensure balanced errors across groups.