Mohammad Taha Askari

At UBC’s Data Communications Group, I conduct graduate research on probabilistic modeling and sequence design for communication systems with memory and nonlinear dynamics. I developed an analytical low-pass filter model to characterize how probabilistic shaping interacts with nonlinear channels and used it to guide interpretable system-level design.

I also proposed a sequence selection algorithm that improves robustness to nonlinear interference by exploiting temporal dependencies, and designed a Bayesian carrier phase recovery method that incorporates prior distributions for improved low-SNR performance. Building on these foundations, I developed a series of neural shaping frameworks, from LSTM-based joint-distribution learning to a Transformer-based sequential architecture with rate-loss-aware training, that achieve state-of-the-art performance on nonlinear fiber channels while remaining compatible with practical PAS/FEC transceivers.

At Nokia Bell Labs, I led ML research on neural sequence modeling for joint distribution learning in optical fiber systems. I designed and implemented Neural Probabilistic Shaping (NPS), an autoregressive LSTM framework trained end-to-end via Gumbel-Softmax sampling, achieving substantial AIR gains over marginal-distribution shaping and sequence selection.

I extended NPS to Neural Probabilistic Amplitude Shaping (NPAS), constraining learning to unsigned symbols for full compatibility with practical PAS/FEC transceivers. I then developed Sequential NPAS (Seq-NPAS), a block-less Transformer-based encoder with a rate-loss-aware training objective; the first neural shaping method to outperform all baselines after accounting for all implementation losses.

At Roche Canada, I worked in an applied ML and software engineering environment on large-scale sequencing data. I built pipelines for feature extraction, dataset generation, and neural network training to support anomaly detection workflows.

I developed and optimized weakly supervised deep learning models, including a multi-label architecture for multi-class anomaly detection and segmentation. I also integrated model evaluation, interpretability, and computational complexity analysis to support iterative improvement.