Abdul Wahab

wahab1 (at) ualberta (dot) ca

I am a graduate student working with Dr. Martha White at the University of Alberta. I work on exploration and representation learning for reinforcement learning.

Previously, I worked as a machine learning engineer at Veeve.io, with Dr. Faisal Shafait and Dr. Ahmad Salman. I developed vision-based solutions for different modules of a smart shopping cart, like cart-state tracking, hand-motion analysis, trajectory analysis, gesture prediction, shrinkage control, item classification, and barcode-through-vision.

Before that I did my undergraduate from the National University of Sciences & Technology (NUST), Islamabad, Pakistan, working in TUKL-NUST R&D Lab with Dr. Faisal Shafait, and Dr. Arsalan Ahmad. I have had the pleasure of visiting the Augmented Vision lab DFKI Kaiserslautern, Germany, as a research associate intern to work on analyzing line-scan camera inputs, with Dr. Gerd Reis and Dr. Didier Stricker.

CV  /  LinkedIn  /  Google Scholar  /  Github

My research interests include:

• Exploration in reinforcement learning
• Representation learning
• Continual learning

In this work, we aim to combine theoretically backed exploration algorithms with sample-efficient architectures.

In this project, we use Two-timescale networks (TTN) in the Offline-Online setting, in which an agent is trained on offline data, and is then allowed to update the representation and the policy networks online. We empirically show that TTN is well suited for the Offline-Online setting as the online updates are more stable and the model converges relatively faster. We also study, analyze and compare different representation learning approaches like sparsity (FTA), input transformations (sparse, random, random-sparse), augmentation, and self-supervised contrastive losses in different combinations, to identify the best combination of these representation learning approaches.

We propose a general method based on bi-directional sequential models (LSTMs) for short-term energy load forecasting suitable for individual household energy forecasting and aggregated energy forecasting for a region.

We propose and compare algorithms to allocate upstream PON capacity, where multiple virtual operators generate independent frame-level allocation over a shared infrastructure. Our fragmentation-aware approach shows the ability to limit latency increase to a few microseconds.

In this project, our goal is to couple combinatorial game theory (CGT) enhancements such as endgame database, and subgame simplifications with the Monte-Carlo Tree Search (MCTS) Solver to design a strong 2D-Clobber player.

Human visual system depends on imagery representations in addition to the visual input to build a concrete holistic picture of the object of interest. These representations are correlated and shared in the human brain and help provide a crisp understanding of the environment. In this work, we study if a similar correlation exists in Artificial Neural Networks (ANNs) and if the correlation is strong enough to use the corresponding fMRI data of both perceptual and quasi-perceptual experiences interchangeably for downstream tasks such as object category prediction.

In this work, we explored the advantages of using a line-scan camera in industrial applications, where the point of observation is a single plane, for monitoring, classification, and segmentation of moving objects. A ConvLSTM-based architecture was proposed, which extracts temporal-correlation from the sequence of line-scans to compensate for the loss of spatial-information in the individual line-scan (single row of pixels). The methodology evaluated on road-crack segmentation datasets outperforms the current benchmark on four public datasets. The methodology was also evaluated on a novel dataset collected on the roads of Islamabad, Pakistan. The methodology was also implemented for a seed-sorting system.

Abdul Wahab. Design inspired by this website.