Hamid Kamkari
 |
Machine Learning Research Intern at Master of Science at Bachelor of Science at Contact: hamidrezakamkari (at) gmail (dot) com |
About Me
Delving into the nexus of machine learning fundamentals, focusing on AI reliability, explainability, and their ties to statistical foundations.
Currently, I am interning at Layer6 AI as a Machine Learning Scientist, delving deep into the reliability of generative models. Specifically, I am looking at pathological behaviours that modern generative models exhibit from a Manifold Learning Theory perspective when they are employed on Out-of-Distribution data in the wild.
As a competitive programmer passionate about computer science theory and practice, I prioritize integrating top engineering practices. Recognizing the dynamic nature of machine learning research, I create libraries in my free time to support researchers and work on projects aimed at bridging the gap between abstract concepts of machine learning with real-world applications in computer vision, healthcare, and computational biology.
Highlights
-
Explaining a Popular Paradox in Deep Generative Models September 2023Layer6 AI , University of Toronto - Toronto, Canada
-
Causal Discovery and Inference using Normalizing Flows May 2023Vector Institute, University of Toronto - Toronto, Canada
-
Dysweep: Enhanced Sweeps for Systematic Experimentation January 2023Vector Institute - Toronto, Canada
-
Attention-based Drug Discovery June 2022Sharif University - Tehran, Iran
-
RNA Sequence design using Graph Neural Networks September 2021Aalto University - Espoo, Finland
-
Semi-definite Programming using Slime Molds June 2021Max-Planck for Informatics - Saarbrücken, Germany
We examine the paradox of deep generative models assigning high likelihoods to unseen data and propose a method to improve their reliability and theoretical understanding in out-of-distribution detection.
Introduced a novel neural network architecture that can learn to understand the underlying structure of the data-generating process. This ultimately helps us produce reliable and explainable models that can account for interventional distributions unseen in the training data. [Code] [Paper]
An integration with Weights & Biases that provides a pipeline to aid reproducibility, continuous development, and large-scale benchmarking. [Code]
Thesis leveraged attention mechanisms in deep learning to identify synergistic drug combinations for cancer research. Achieved a significant accuracy boost of 10% in dose response prediction for the NCI-ALMANAC cancer drug database.
Designing beneficial RNA structures is challenging in biotechnology. We use reinforcement learning algorithms combined with graph neural networks to model and design RNA sequences, obtaining previously underexplored structures like RNA pseudo-knots.
Inspired by the optimization dynamics of slime molds in nature, developed a mathematical dynamic that provably converges to the optimal solution for semi-definite programming problems. [Paper]