Mohammad Taha Askari

My research has focused on probabilistic signal shaping methods, particularly probabilistic amplitude shaping for optical fiber communications. I introduced a low pass filter model to explain the complex interplay between probabilistic shaping and fiber nonlinearity. Building on this theoretical work, I developed a novel sequence selection scheme to improve nonlinearity tolerance in probabilistic amplitude shaping for optical fiber communications.

More recently, I designed a Bayesian phase search algorithm for carrier phase recovery in the low-SNR regime, addressing specific challenges that arise in probabilistic shaping scenarios. My current work explores end-to-end sequence-based auto encoder approaches for probabilistic shaping, leveraging advances in deep learning for communication systems.

During my internship at Roche Canada, I designed and implemented comprehensive pipelines for feature extraction, dataset generation, and neural network training. These systems enabled efficient anomaly detection in complex sequencing data. I developed and optimized weakly supervised neural network models, including a novel multi-label architecture specifically designed for multi-class anomaly segmentation.

A key contribution of my work was integrating interpretability and complexity analysis into the training workflow. This allowed me to deliver detailed reports and insights that facilitated model evaluation and enhancement, making the technology more accessible to non-technical stakeholders and improving overall performance.