Yogesh Kumar

Education

Jan 2019 – Ongoing PhD Researcher, Aalto University, Espoo, Finland

Sample-Efficient Training of Large Neural Networks - Optimized large model training in data-scarce contexts using transfer learning, improving performance in tasks like generating radiologist reports with RAG and fine-tuning for EHR data.
Enhancing Model Interpretability - Refined the Centered Kernel Alignment (CKA) method to improve neural network similarity measurements, addressing confounders to better reveal functional similarities across domains.
Large-Scale Healthcare Utilization Predictions - Developed SANSformer, an attention-free sequential model tailored for EHR, improving healthcare utilization predictions with large-scale datasets.

Apr 2022 – May 2023 Senior AI/ML Software Engineer, Nokia Platforms, Espoo, Finland

Played a key role in building Nokia's proprietary AI/ML engine, managing the model development lifecycle.
Architected and deployed a multi-GPU training pipeline, enhancing model training speed and efficiency.
Designed custom CNN architectures for digital receivers, optimizing for hardware constraints and edge devices.
Introduced scrum methodologies, improving team workflows and collaboration, and served as a stand-in Local Product Owner.

Sep 2017 – Mar 2018 Data Scientist, Curefit, Bangalore, India

Led development of the Virtual Trainer MVP, using computer vision to provide real-time workout pose corrections.
Developed an RNN model to analyze phone accelerometer data, achieving 87% accuracy in sleep pattern detection.

Jan 2017 – Aug 2017 Data Scientist, Revmax, New York, NY

Led data science initiatives for an early-stage startup, enhancing routing efficiency and vehicle utilization for New York City cabs.

Feb 2016 – May 2016 Technical Specialist, CITI Digital Innovation Lab, New York, NY

Deployed an ELK framework to aggregate, parse, and visualize server logs from distributed infrastructure.

Aug 2013 – Aug 2015 Software Engineer, ISO New England, Holyoke, Massachusetts

Contributed to incremental Scrum releases for a Java EE web application, resulting in a 73% surge in web traffic.
Integrated JavaScript/Angular and Ajax-driven elements to enhance UI responsiveness.

Improving Medical Multi-modal Contrastive Learning with Expert Annotations - ECCV 2024
Self-Supervised Forecasting in Electronic Health Records with Attention-Free Models - IEEE Transactions on AI (ML4H NeurIPS 2023 Best Findings Paper)
Deconfounded Representation Similarity for Comparison of Neural Networks - NeurIPS 2022 (Oral Presentation)
Predicting Utilization of Healthcare Services from Disease Trajectories using RNNs - ML4H at NeurIPS 2019

NASA Pose Bowl: Spacecraft Detection Track - DrivenData 2024 (Rank 10 out of 651 teams)
PANDA Challenge - Kaggle 2020 (Top 3% of 1010 teams)
Generative Dog Images - Kaggle 2019 (Top 3% of 927 teams)