Machine Learning Engineer
Resume:
POOYA TAVALLALI
PROFESSIONAL EXPERIENCE
Machine Learning Engineer
Iterative Health Boston, MA
January 2022 - Present
● 2022:
Introduced machine learning techniques for video model interpretability.
Delivered interpretable ML models meeting product requirements for production deployment.
Developed a reliable evaluation technique for model interpretation.
Conducted data and annotation quality analysis and visualization.
Designed and trained a model for ulceration level classification in colonoscopy videos using transfer learning. Used for Crohn’s Disease scoring.
Implemented data ingestion to databases.
Created annotation policy for landmark segmentation in colonoscopy videos.
Trained an ensemble of deep convolutional neural networks for landmark segmentation, achieving human-level performance.
Developed a data ingestion pipeline and ensured QA of ML pipelines.
Designed systems for ML, focusing on annotation quality analysis.
● 2023:
Focused on ML infrastructure and pipeline QA, including data annotation analytics.
Enhanced ML dataset creation and training pipeline for Cecum classification
■ Image sequences and aiming to find Cecum images as the start of the procedure
■ Designed evaluation metrics given the nature of the problem
■ Model achieved 88.1% accuracy in finding the first Cecum image as the start of procedure
■ designed an optimization for automated labeling of missing labels.
● Improved accuracy to 89.8%
Contributed to ML pipelines system design and onboarding for new ML members.
Improved ML pipelines for polyp detection, measuring performance metrics for detection models based on FDA requested metrics.
● 2024:
Designed an evaluation pipeline for trained polyp detection model evaluation
■ The evaluation pipeline was designed to run at scale and with low cost (reduced the cost of using existing pipeline by 95%, used for model parameter tuning)
Maintained, improved and developed ML pipelines
Trained multiple polyp detection models using recent detection models
■ Could outperform existing polyp detection model
■ Model was several times smaller than original model (useful for faster and cheaper inference)
Optimized model for edge deployment (performance and cost improved)
Trained multiple classification and detection models that worked together for improved detection Graduate Research Assistant
University of California, Merced (PhD) 2016 – 2021
● Developed EM-based neural networks simulating prototype nearest neighbors and optimized adversarial attack/defense strategies for multi-class models.
● Conducted epidemiological research on air pollution's impact on emergency visits using case-crossover studies.
● Real-time Face detection using Viola-Jones face detector
● Published research on topics interpretable models, decision trees, and adversarial robustness (ICIP, NeurIPS). Key Projects
Regions of Interest (2022):
● Developed a model to locate critical time intervals within videos that contribute to final predictions, ensuring interpretability with statistically significant results. Crohn’s Disease Scoring (2022):
● Designed a deep learning model for ulceration classification using transfer learning and k-fold cross-validation to handle small datasets.
Landmark Detection in Colonoscopy Videos (2022):
● Created annotation policies and trained an ensemble of CNNs for temporal modeling. Achieved human-level accuracy in identifying anatomical landmarks. Cecum Classification Pipeline (2023):
● Solved an optimization problem to label missing data in sequential images. Final model outperformed trained human annotators.
Polyp Detection (2023-2024):
● Designed and implemented a polyp detection model using limited data with mixed annotations
(bounding boxes and classification labels).
● Created an evaluation framework to measure detection performance at both frame and video levels. Skills
Programming Languages: Python, C++
Frameworks and Libraries: Pytorch, TensorFlow, Keras, Numpy, Scikit-Learn, Scipy, Pandas, SQL, GitHub Statistical Software: Matlab, R.
Cloud-Based Services: AWS, GCP
Technical Skills and Interests
Core Competencies: Computer Vision, Machine Learning Pipelines, Deep Learning Techniques, Data Handling & Analysis, Cloud Services & DevOps (AWS), Optimization & Interpretability Topics of Interest: Machine Learning, Deep Learning, Computer Vision, Large Language Models, Image Classification, Video Processing, Data Analysis, Statistical Learning, Model Interpretability, Data Clustering, Tracking, Applications of Interpretable Machine Learning in Biostatistics, Optimization, Machine Learning Programming in Python
Education
[2021] PhD -Electrical Engineering and Computer Science, UNIVERSITY OF CALIFORNIA MERCED, USA
[2016] MSc - Electrical Engineering, SHIRAZ UNIVERSITY, IRAN
[2013] BSc - Electrical Engineering, SHIRAZ UNIVERSITY, IRAN Selected Papers
[2021] “Adversarial Poisoning Attacks and Defense for General Multi-Class Models Based On Synthetic Reduced Nearest Neighbors”, Pooya Tavallali, et. al; International Conference on Image Processing (ICIP)
[2020] “Interpretable Synthetic Reduced Nearest Neighbor: an Expectation Maximization Approach”, Pooya Tavallali, Peyman Tavallali, Mohammad Reza Khosravi and Mukesh Singhal; International Conference on Image Processing (ICIP)
[2019] “Robust Cascaded Skin Detector Based on Adaboost”, Pooya Tavallali, Mehran Yazdi, and Mohammad Reza Khosravi; Multimedia Tools and Applications, pages 1–22
[2018] “Alternating Optimization of Decision Trees, with Application to Learning Sparse Oblique Trees”, Miguel Á. CarreiraPerpiñán and Pooya Tavallali; In Advances in Neural Information Processing Systems(NeurIPS)
Contact this candidate