Aamer Abdul

Actively looking for ML internship roles for summer 2024.

I am a Masters student at Mila Quebec where I am advised by Professor Samira Ebrahimi Kahou, Professor Rita Noumeir and Dr. Philippe Jouvet. Additionally, I collaborate with CHU Sainte-Justine hospital as a graduate researcher. My research is broadly in computer vision, machine learning and its applications to healthcare. In particular, I work on self supervised learning for videos and the application of offline reinforcement learning to safety-critical environments such as healthcare.

Previously, I was a Research Intern at the University of Toronto, working under the supervision of Professor Chi-Guhn Lee, where I worked on adversarial domain adaptation for data-sparse image classification. My undergrad was in Electronics and Communication Engineering from BITS Pilani, during which I mainly worked on computer vision and machine listening.

Email  /  Resume  /  Publications  /  Google Scholar  /  Twitter  /  LinkedIn

• October 2023: Spotlight Paper @ NeurIPS Workshop - Our work Empowering Clinicians with MeDT: A Framework for Sepsis Treatment was accepted at NeurIPS 2023 Workshop on Goal-Conditioned Reinforcement Learning.
• July 2023: Paper @ Preprint release - Preprint release of our work Transformers in Reinforcement Learning: A Survey.
• May 2023: Reviewer - Served as reviewer for ICCV.
• November 2022: Paper @ NeurIPS Workshop - Our work Learning from Uncertain Concepts via Test Time Interventions was accepted at NeurIPS 2022 Workshop on Trustworthy and Socially Responsible Machine Learning.
• November 2022: Paper @ NeurIPS Workshop - Our work Pitfalls of Conditional Batch Normalization for Contextual Multi-Modal Learning was accepted at ICBINB NeurIPS 2022 Workshop .
• June 2022: Paper @ CVPR - Our work Revisiting learnable affines for batch norm in few-shot transfer learning was accepted at CVPR 2022.
• May 2022: Position - Started Masters by Research at Mila, in collaboration with CHU Sainte-Justine.
• Dec 2021: Paper @ NeurIPS Workshop - Our work Shift and Scale is Detrimental To Few-Shot Transfer was accepted at NeurIPS 2021 Workshop on Distribution Shifts: Connecting Methods and Applications.
• July 2021: Position - Started Research Internship at Mila.
• May 2021: Paper - Our work Robust segmentation of vascular network using deeply cascaded AReN-UNet was accepted to the Biomedical Signal Processing and Control journal.
• Dec 2020: Position - Started Research Internship at University of Toronto.

Offline reinforcement learning is promising for safety-critical tasks like clinical decision support, but faces challenges of interpretability and clinician interactivity. To overcome these, the proposed Medical Decision Transformer (MeDT) utilizes a goal-conditioned RL paradigm for sepsis treatment recommendations. MeDT employs the decision transformer architecture, considering factors like treatment outcomes, patient acuity scores, dosages, and current/past medical states to provide a holistic view of the patient's history. This enhances decision-making by allowing MeDT to generate actions based on user-specified goals, ensuring clinician interactability and addressing sparse rewards. Results from the MIMIC-III dataset demonstrate MeDT's effectiveness in producing interventions that either outperform or compete with existing methods, offering a more interpretable, personalized, and clinician-directed approach.

We present a comprehensive survey on the emergence of transformers in reinforcement learning.

We introduce a novel normalization strategy for few-shot transfer learning, bringing performance improvements upto 5% in accuracy with neglible computational overhead.

We demonstrate that using conditional batch normalization for contextual multimodal tasks leads to shortcut learning. Our experiments reveal that the network learns close to no visual features and relies solely on contextual auxiliary data to make predictions.

We propose a novel uncertainty based strategy to select the interventions in Concept Bottleneck Models. We demonstrate the proposed method can improve human-model interaction via more meaningful corrections, while reducing information leakage.

We present a U-Net segmentation architecture based on cascading and attention modules for improved performance on retinal blood vessel segmentation.

A study on effectiveness of data augmentation strategies on downstream performance for urban sound classification with CNNs.

An object detection approach with Faster RCNN's to detect and track locust insects.

The template is borrowed from Jon Barron's homepage