Mihkel Lepson

University of Tartu
1 previous affiliation
    • PhD student · Since 2023
    • MS student · 2020–2023
    • Undergrad student · 2017–2020
3Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Mihkel Lepson focuses on the development and evaluation of machine learning models, with specific applications in trajectory forecasting and medical natural language processing. His research addresses the mathematical rigor of evaluation metrics for multimodal predictions and the utility of synthetic data for training models in low-resource linguistic and clinical contexts. Lepson’s recent work emphasizes the transition from traditional heuristic-based evaluation to distribution-aware frameworks and the application of large language models (LLMs) to specialized domains where annotated data is scarce.

Research Trajectory

Lepson’s research trajectory is characterized by a dual focus on the foundational reliability of machine learning evaluations and the practical deployment of models in high-stakes, data-constrained environments. In 2024, his work concentrated on the problem of trajectory prediction, a critical component in autonomous systems and robotics. This research identified a significant gap in how the field measures success: while modern models are increasingly designed to be multimodal—predicting a range of possible future paths—the metrics used to evaluate them often fail to account for the quality of the entire predicted distribution.

Parallel to this work in predictive modeling, Lepson has contributed to the field of clinical natural language processing (NLP). Starting in 2024 and continuing into 2025, this research direction has focused on the challenges of Named Entity Recognition (NER) for low-resource languages, specifically Estonian. This work addresses the intersection of data privacy and model performance, exploring how synthetic data generated by LLMs can serve as a viable substitute for sensitive human-annotated medical records. This shift represents an evolution from purely algorithmic evaluation toward the development of end-to-end pipelines for specialized domain adaptation.

Current Directions

Lepson’s current research efforts are centered on two primary themes: the refinement of scoring rules for probabilistic forecasting and the validation of synthetic data for clinical applications.

In the domain of trajectory forecasting, Lepson has worked on establishing more robust benchmarks for multimodal models. Traditional metrics, such as "Minimum of N" (MoN), have been criticized for being "improper" scoring rules, meaning they can be optimized by models that do not accurately represent the true underlying probability distribution. Lepson’s work advocates for the adoption of strictly proper scoring rules, which ensure that the highest scores are only achieved by models that correctly estimate both the mean and the variance of future trajectories.

In the area of medical NLP, Lepson is investigating the scalability of LLM-based synthetic data generation. The primary objective is to facilitate the training of robust NER models for languages like Estonian, which lack the massive annotated corpora available for English. By leveraging the few-shot capabilities of LLMs to produce high-quality synthetic clinical text, Lepson and his collaborators aim to develop models that can extract critical medical entities—such as diagnoses, medications, and procedures—without compromising patient confidentiality. This work involves rigorous validation to ensure that models trained on synthetic data generalize effectively to real-world clinical documentation.

Representative Work

In Evaluation of Trajectory Distribution Predictions with Energy Score, Lepson and his co-authors critique the standard practices in trajectory prediction evaluation. The paper demonstrates that the widely used "Minimum of N" (MoN) metrics—which measure the error of the single best prediction out of a set of samples—are mathematically improper. The authors show that these metrics can reward models that produce biased or over-dispersed distributions. As an alternative, the paper proposes the use of the Energy Score, a strictly proper scoring rule that evaluates the entire predicted distribution against the ground truth. This work provides a more reliable framework for assessing the performance of multimodal forecasting models in fields such as autonomous driving.

Lepson’s contributions to clinical informatics are detailed in Using Synthetic Health Care Data to Leverage Large Language Models for Named Entity Recognition: Development and Validation Study. This research addresses the scarcity of annotated medical data for the Estonian language. Lepson participated in the development of a novel approach that uses LLMs to generate synthetic healthcare records, which are then used to train NER models. The study validates this method by comparing the performance of models trained on synthetic data against those trained on limited human-annotated sets. The results suggest that synthetic data can effectively bridge the gap in low-resource settings while maintaining a high degree of privacy, as the training process avoids direct exposure to sensitive patient information. This methodology offers a potential blueprint for expanding clinical NLP capabilities to other underrepresented languages and specialized medical domains.

Given the recent nature of Lepson's documented research (2024–2025), his work currently represents an emerging contribution to the fields of probabilistic evaluation and synthetic data utility. The available evidence highlights a consistent focus on improving the reliability and accessibility of machine learning tools in complex, real-world scenarios.

Profile updated

See something wrong? Report a mistake