Karl Pichotta

Karl Pichotta

Memorial Sloan Kettering Cancer Center
3 previous affiliations
3Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Karl Pichotta focuses on natural language processing and machine learning, with specific interests in ordinal prediction, script knowledge, and privacy in clinical language models. His trajectory spans from developing robust discretization techniques for continuous data modeling to evaluating the security of large language models in healthcare. Recent work investigates the privacy implications of pretraining models on sensitive clinical notes.

Research Trajectory

The available evidence outlines a research trajectory that bridges core machine learning methodology with applied natural language processing. In 2018, Pichotta contributed to methodological improvements in conditional probability estimation. As second author on Leaf-Smoothed Hierarchical Softmax for Ordinal Prediction, he helped develop a specialized hierarchical softmax variant inspired by k-d trees. This work leveraged the spatial structure of multivariate ordinal labels and introduced a novel regularizer based on signal processing on noisy graphs to improve the sample efficiency of deep learning models. This research aligns with broader research directions such as Discretization for Robust and Efficient Continuous Data Modeling and the area of Non-parametric Density Estimation via Target Discretization.

Moving into natural language processing, Pichotta participated as a middle author in collaborative projects evaluating language models and structured knowledge. In 2019, he contributed to Benchmarking Hierarchical Script Knowledge, which explored the representation and evaluation of script knowledge in NLP systems. This work touches upon themes related to Language Independent NLP without Linguistic Pipelines, reflecting an interest in how models capture complex, structured sequences of events.

The corpus indicates a subsequent shift toward the intersection of clinical informatics and model privacy. In 2021, Pichotta was a co-author on Does BERT Pretrained on Clinical Notes Reveal Sensitive Data?. This research investigated whether large language models trained on sensitive healthcare data could inadvertently leak protected health information, highlighting the growing need for rigorous privacy evaluations in applied NLP.

Current Directions

The most recent paper in the evidence, from 2021, addresses the privacy and security of large language models in specialized domains. Specifically, this direction investigates the vulnerabilities of models like BERT when pretrained on sensitive datasets, such as clinical notes. This suggests a focus on understanding and mitigating the risks of data leakage in applied natural language processing, ensuring that models can be safely evaluated and deployed in high-stakes environments like healthcare.

Profile updated

See something wrong? Report a mistake