About Me
I am an AI researcher working on making deep learning efficient when compute, data, or communication is the bottleneck: fitting capable models into the memory of a single device, training them on less data, and coordinating them across many devices without moving the data.
I am a PhD candidate in Electrical Engineering at KAIST, working with Prof. Jaekyun Moon in the Moon Lab. My dissertation is on model merging for foundation-model experts: combining and pruning specialized models so that many of them can be stored, served, and updated within the memory budget of a single device, without retraining from scratch. Along the way I have worked on communication-efficient federated and split learning (EvoFed, NeurIPS 2023), pruning-guided data selection for efficient training (PruneFuse, TMLR Featured Certification, which cuts selection-phase compute by up to 96%), and decoupled, layer-wise training of deep networks.
More recently I have been focusing on Voice AI problems, in particular the evaluation of text-to-speech and speech-synthesis systems. Neural audio generation still fails in ways that transcript- and preference-based metrics such as WER and MOS do not register: word-level pronunciation errors, mishandled text normalization, and drift in speaker identity. I am interested in benchmarks and evaluation methods that make these failures visible and measurable.
Before KAIST, I earned a B.Sc. in Electrical Engineering from the University of Engineering and Technology (UET), Lahore, where I was fortunate to work with Dr. Muhammad Tahir.
Research Interests
- Voice AI: evaluation of text-to-speech and speech-synthesis systems; pronunciation fidelity and other failure modes of neural audio generation that WER- and MOS-style metrics cannot observe.
- Model Merging and Sparse Experts: data-free merging of pruned and specialized foundation-model experts; spectral-sparse decomposition; storage- and memory-efficient expert serving.
- Efficient and On-Device ML: network pruning, model compression, and data-efficient training (pruning-guided active learning) for resource-constrained deployment.
- Distributed and Federated Learning: communication-efficient federated and split learning under heterogeneous clients.
News
- [Sep. 2026] Serving as a reviewer for NeurIPS 2026, ICLR 2027, and AAAI 2027, and for the GlobalSouthAI (NeurIPS 2026) and GlobalSouthML (ICML 2026) workshops.
- [July 2026] Presented “PruneFuse: Efficient Data Selection via Weight Pruning and Network Fusion” at ICML 2026 in Seoul, South Korea.
- [Mar. 2026] Our paper, “PruneFuse: Efficient Data Selection via Weight Pruning and Network Fusion,” received a Featured Certification (Spotlight) in Transactions on Machine Learning Research (TMLR).
- [Oct. 2024] Our paper, “Pruning-based Data Selection and Network Fusion for Efficient Deep Learning,” was accepted to the Attributing Model Behavior at Scale Workshop at NeurIPS 2024.
- [Oct. 2023] Our paper, “EvoFed: Leveraging Evolutionary Strategies for Communication-Efficient Federated Learning,” was accepted to NeurIPS 2023.
Publications
-
Merging in the Small Room: A Spectral-Sparse Decomposition for Pruned ExpertsUnder review, 2026.A spectral-sparse decomposition for merging pruned expert models without any data, so many specialised models can be stored and served at a fraction of the cost.Under review -
TMLR
Transactions on Machine Learning Research (TMLR), Mar 2026. Presented at ICML 2026.Combines structured pruning with weight-aligned model merging to warm-start training after data selection, cutting selection-phase compute by up to 96%. -
NeurIPS W
Workshop on Attributing Model Behavior at Scale (ATTRIB) at NeurIPS, 2024.Early version of PruneFuse; selects data with a pruned surrogate network and fuses its weights into the target model to avoid training from scratch. -
NeurIPS
Advances in Neural Information Processing Systems (NeurIPS), 2023.Clients send compact fitness scores against a shared population of noise-perturbed models instead of model parameters, matching FedAvg accuracy while cutting communication drastically. -
ICML W
Hardware Aware Efficient Training (HAET) Workshop at ICML, 2022.Periodically reinstates the global objective into local-loss, layer-wise training (PGL), recovering the generalisation that decoupled training normally loses while keeping its low memory footprint. -
IC2S2
International Conference on Computational Social Science (IC2S2), 2022 and Korea Computer Congress (KCC), 2022.Uses large-scale web data to trace how gender pay gaps in STEM careers emerge and widen across life trajectories. -
ICML W
Workshop on Federated Learning for User Privacy and Data Confidentiality at ICML, 2021.Lets clients in federated split learning update from locally generated losses, cutting round-trip communication with the server.
Honors and Awards
- Featured Certification (Spotlight), Transactions on Machine Learning Research, for PruneFuse, 2026.
- KAIST International Students Representative, KAIST, 2023–2026.
- EE International Office Ambassador, School of Electrical Engineering, KAIST, 2023–2024.
- Young Future Energy Leader, Khalifa University, 2022–2023.
- Double Gold Medalist, B.Sc. Electrical Engineering, UET Lahore, 2020.
- Dean’s Roll of Honor, UET Lahore — five semesters, 2016–2019.
- Best Microprocessor Project Award, Department of Electrical Engineering, UET Lahore, 2018.
- Pakistan Engineering Congress Scholarship, 2018 and 2019.
Teaching
- Teaching Assistant, Machine Learning Basics and Practices — School of Electrical Engineering, KAIST, Spring 2025.
- Teaching Assistant, Introduction to Programming — Department of Electrical Engineering, UET Lahore, Fall 2019.
- Teaching Assistant, Digital Systems — Department of Electrical Engineering, UET Lahore, Fall 2018.
Services
Conference Reviewer
- International Conference on Learning Representations (ICLR) 2025, 2026, 2027
- International Conference on Machine Learning (ICML) 2025
- Artificial Intelligence and Statistics (AISTATS) 2025, 2026
- Advances in Neural Information Processing Systems (NeurIPS) 2024, 2025, 2026
- AAAI Conference on Artificial Intelligence (AAAI) 2027
Workshop Reviewer
- Global South in AI (GlobalSouthAI) workshop at NeurIPS 2026
- Global South in ML (GlobalSouthML) workshop at ICML 2026