Muhammad Usama

Muhammad Usama

AI Initiative Team Lead, Braindeck Inc. | Ph.D., KAIST

Publications

You can also find my articles on my Google Scholar profile.

Journal Articles

The GRPO Tax is Smaller Than You Think: A Longitudinal Study of Capability Preservation During Reasoning Training

Published in Transactions on Machine Learning Research (TMLR), 2026

A dense-checkpoint longitudinal study of how much general capability language models retain when trained with GRPO for reasoning.

M. Usama, "The GRPO Tax is Smaller Than You Think: A Longitudinal Study of Capability Preservation During Reasoning Training," Transactions on Machine Learning Research (TMLR), 2026.
Paper · Code

Predicting Pediatric Myopia Progression from Single Screening Encounters: A Multi-Outcome Development and Validation Study

Published in Journal of Translational Medicine, 2026

A multi-outcome model for predicting pediatric myopia progression from a single screening encounter, developed and validated on longitudinal screening data.

N. Wei, C. Li, S. Moutari, M. Usama, J. Li, Q. Chen, E. E. Pazo, and X. Qian, "Predicting Pediatric Myopia Progression from Single Screening Encounters: A Multi-Outcome Development and Validation Study," Journal of Translational Medicine, 2026. doi: 10.1186/s12967-026-09003-2.
Paper

Machine Learning-Based Prediction of Cycloplegic Refraction Using the Eyerobo Vision Screener: Design—A Cross-Sectional Comparative Device Study

Published in Ophthalmology and Therapy, 2026

A cross-sectional device study evaluating machine learning models for predicting cycloplegic refraction from a portable vision screener.

M. Xia, B. Yao, Y. Wang, F. Liang, W. Zhao, L. Qin, S. Liu, M. Usama, Z. Zhang, W. Guo, E. E. Pazo, and X. Ren, "Machine Learning-Based Prediction of Cycloplegic Refraction Using the Eyerobo Vision Screener: Design—A Cross-Sectional Comparative Device Study," Ophthalmology and Therapy, 2026. doi: 10.1007/s40123-026-01482-2.
Paper

Distributional Reinforcement Learning with Information Bottleneck for Uncertainty-Aware DRAM Equalization

Published in IEEE Transactions on Components, Packaging and Manufacturing Technology, 2026

A distributional reinforcement learning approach, regularized with an information bottleneck, for optimizing DRAM equalizer parameters under signal uncertainty.

M. Usama and D. E. Chang, "Distributional Reinforcement Learning with Information Bottleneck for Uncertainty-Aware DRAM Equalization," IEEE Transactions on Components, Packaging and Manufacturing Technology, vol. 16, no. 8, pp. 1912–1929, 2026. doi: 10.1109/TCPMT.2026.3674506.
Paper

Validation of the Eyerobo FC Portable Fundus Camera for Diabetic Retinopathy Screening Using Public Datasets and Deep Learning

Published in Ophthalmology and Therapy, 2026

Validation of a portable fundus camera for diabetic retinopathy screening against public retinal imaging datasets using deep learning models.

E. E. Pazo, X. Liu, S. Liu, Q. Zhang, M. Zhang, Z. Zhang, S. Moutari, M. Usama, and X. Ren, "Validation of the Eyerobo FC Portable Fundus Camera for Diabetic Retinopathy Screening Using Public Datasets and Deep Learning," Ophthalmology and Therapy, 2026. doi: 10.1007/s40123-026-01353-w.
Paper

Multi-Lane Joint DRAM Equalization via Cooperative Multi-Agent Distributional Reinforcement Learning

Published in IEEE Transactions on Components, Packaging and Manufacturing Technology, 2026

A cooperative multi-agent distributional reinforcement learning framework for jointly optimizing equalizer parameters across multiple DRAM lanes.

M. Usama and D. E. Chang, "Multi-Lane Joint DRAM Equalization via Cooperative Multi-Agent Distributional Reinforcement Learning," IEEE Transactions on Components, Packaging and Manufacturing Technology, 2026. doi: 10.1109/TCPMT.2026.3711630.
Paper

Memory-Augmented Architecture for Long-Term Context Handling in Large Language Models

Published in IEEE Transactions on Artificial Intelligence, 2026

A memory-augmented architecture with dynamic retrieval, update, and pruning mechanisms for maintaining coherence over long dialogues in large language models.

H. U. K. Shinwari and M. Usama (equal contribution), "Memory-Augmented Architecture for Long-Term Context Handling in Large Language Models," IEEE Transactions on Artificial Intelligence, 2026. doi: 10.1109/TAI.2026.3701680.
Paper

Structure-Preserving Super-Resolution of Retinal Fundus Images via a Dual-Transformer Residual Network

Published in Frontiers in Medicine (Section: Ophthalmology), 2025

A dual-transformer residual network for super-resolving retinal fundus images while preserving diagnostically relevant structure.

E. E. Pazo, S. Moutari, F. Gao, L. Hu, M. Usama, X. Li, and J. Liu, "Structure-Preserving Super-Resolution of Retinal Fundus Images via a Dual-Transformer Residual Network," Frontiers in Medicine (Section: Ophthalmology), 2025. doi: 10.3389/fmed.2025.1730678.
Paper

Deep Reinforcement Learning-Based DFE Structure Optimization for High-Speed DRAM Signals

Published in IEEE Transactions on Components, Packaging and Manufacturing Technology, 2025

A deep reinforcement learning method for optimizing decision feedback equalizer (DFE) structure for high-speed DRAM signal integrity.

M. Usama, H.-D. Jang, and D. E. Chang, "Deep Reinforcement Learning-Based DFE Structure Optimization for High-Speed DRAM Signals," IEEE Transactions on Components, Packaging and Manufacturing Technology, 2025. doi: 10.1109/TCPMT.2025.3648009.
Paper

ARD-LoRA: Dynamic Rank Allocation for Parameter-Efficient Fine-Tuning of Foundation Models with Heterogeneous Adaptation Needs

Published in IEEE Transactions on Artificial Intelligence, 2025

A dynamic rank allocation method for low-rank adaptation that assigns heterogeneous per-layer ranks based on adaptation needs, evaluated on LLaMA-3.1-70B and PaliGemma-2.

H. U. K. Shinwari and M. Usama (equal contribution), "ARD-LoRA: Dynamic Rank Allocation for Parameter-Efficient Fine-Tuning of Foundation Models with Heterogeneous Adaptation Needs," IEEE Transactions on Artificial Intelligence, 2025. doi: 10.1109/TAI.2025.3605569.
Paper

Learning Flexible and Fair Data Representations

Published in IEEE Access, 2022

A representation learning method for mitigating bias in downstream utility tasks while preserving flexibility for multiple prediction targets.

M. Usama and D. E. Chang, "Learning Flexible and Fair Data Representations," IEEE Access, vol. 10, 2022. doi: 10.1109/ACCESS.2022.3207747.
Paper

Conference Papers

Beyond Adjacent Layers: Graph-Guided Layer Fusion for Compressing Large Language Models

Published in Conference on Neural Information Processing Systems (NeurIPS), 2026

A graph-guided layer-fusion method for compressing large language models that goes beyond merging only adjacent layers.

M. Usama and Y. Jung, "Beyond Adjacent Layers: Graph-Guided Layer Fusion for Compressing Large Language Models," Conference on Neural Information Processing Systems (NeurIPS), 2026.
Code

Uncertainty-Aware DRAM Equalizer Optimization

Published in 2025 Summer Annual Conference of the Korean Institute of Electromagnetic Engineering and Science (KIEES), 2025

An uncertainty-aware optimization approach for DRAM equalizer design.

M. Usama and D. E. Chang, "Uncertainty-Aware DRAM Equalizer Optimization," 2025 Summer Annual Conference of the Korean Institute of Electromagnetic Engineering and Science (KIEES), 2025.

Uncertainty-Aware DRAM Equalizer Optimization Using Bayesian Representation Learning and Deployment Risk Assessment

Published in 2025 Conference on Electronics, Semiconductors, and Artificial Intelligence (CESAI), 2025

A Bayesian representation learning approach to DRAM equalizer optimization with an associated deployment risk assessment.

M. Usama and D. E. Chang, "Uncertainty-Aware DRAM Equalizer Optimization Using Bayesian Representation Learning and Deployment Risk Assessment," 2025 Conference on Electronics, Semiconductors, and Artificial Intelligence (CESAI), 2025.

Representation Learning for Distortion Classification in DRAM Data

Published in 39th Institute of Control, Robotics and Systems (ICROS) Conference, 2024

A representation learning approach for classifying distortion types in DRAM signal data.

M. Usama and D. E. Chang, "Representation Learning for Distortion Classification in DRAM Data," 39th Institute of Control, Robotics and Systems (ICROS) Conference, 2024.

Optimization of Decision Feedback Equalizer Structure through Genetic Algorithm and Neural Network Evaluation

Published in 39th Institute of Control, Robotics and Systems (ICROS) Conference, 2024

A genetic algorithm search over decision feedback equalizer structures, using a neural network for fitness evaluation.

M. Usama and D. E. Chang, "Optimization of Decision Feedback Equalizer Structure through Genetic Algorithm and Neural Network Evaluation," 39th Institute of Control, Robotics and Systems (ICROS) Conference, 2024.

PyEye: An Integrated Approach for Signal Integrity Assessment and Eye Diagram Generation

Published in 2023 IEEE International Conference on Consumer Electronics-Asia (ICCE-Asia), 2023

An integrated software tool for signal integrity assessment and eye-diagram generation.

M. Usama and D. E. Chang, "PyEye: An Integrated Approach for Signal Integrity Assessment and Eye Diagram Generation," 2023 IEEE International Conference on Consumer Electronics-Asia (ICCE-Asia), 2023.

Learning-Based Control of Quadrotor using Gaussian Processes

Published in 6th Joint Conference of the Korean Artificial Intelligence Association (CKAIA 2022), 2022

A Gaussian-process-based learning approach for quadrotor control. Recipient of the Special Society Award at CKAIA 2022.

M. Usama and D. E. Chang, "Learning-Based Control of Quadrotor using Gaussian Processes," 6th Joint Conference of the Korean Artificial Intelligence Association (CKAIA), 2022. Recipient of the Special Society Award.

Learning-Driven Exploration for Reinforcement Learning

Published in 2021 21st International Conference on Control, Automation and Systems (ICCAS), 2021

An entropy-based exploration strategy for reinforcement learning agents.

M. Usama and D. E. Chang, "Learning-Driven Exploration for Reinforcement Learning," 2021 21st International Conference on Control, Automation and Systems (ICCAS), 2021.
Code · Video

Robotic Navigation using Entropy-Based Exploration

Published in 2019 19th International Conference on Control, Automation and Systems (ICCAS), 2019

An entropy-based exploration method applied to robotic navigation.

M. Usama and D. E. Chang, "Robotic Navigation using Entropy-Based Exploration," 2019 19th International Conference on Control, Automation and Systems (ICCAS), 2019.
Video

Control of Furuta Pendulum with Reinforcement Learning

Published in 19th International Conference on Control, Automation, and Systems (ICCAS 2019), 2019

A reinforcement learning controller for the Furuta pendulum swing-up and stabilization task.

D. Lee, M. Usama, and D. E. Chang, "Control of Furuta Pendulum with Reinforcement Learning," 19th International Conference on Control, Automation, and Systems (ICCAS), 2019.
Paper · Video

Towards Robust Neural Networks with Lipschitz Continuity

Published in 17th International Workshop on Digital Forensics and Watermarking (IWDW 2018), 2018

An analysis of Lipschitz continuity as a mechanism for improving the robustness of neural networks.

M. Usama and D. E. Chang, "Towards Robust Neural Networks with Lipschitz Continuity," 17th International Workshop on Digital Forensics and Watermarking (IWDW), 2018.
Paper

Under Review / Submitted

F-ConvNeXt: Frequency-Aware Domain Generalization for Deepfake Detection via Self-Blended Synthetic Training

Under review at Conference on Neural Information Processing Systems (NeurIPS)

A frequency-aware ConvNeXt architecture for domain-generalized deepfake detection, trained with self-blended synthetic data.

M. Usama and Y. Jung, "F-ConvNeXt: Frequency-Aware Domain Generalization for Deepfake Detection via Self-Blended Synthetic Training," Conference on Neural Information Processing Systems (NeurIPS), 2026.

Learning High-Quality Latent Representations for Anomaly Detection and Signal Integrity Enhancement in High-Speed Signals

Under review at IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems

A latent representation learning method for anomaly detection and signal integrity enhancement in high-speed signals.

M. Usama, H.-D. Jang, S. Shanbhag, Y.-C. Sung, S.-J. Bae, and D. E. Chang, "Learning High-Quality Latent Representations for Anomaly Detection and Signal Integrity Enhancement in High-Speed Signals," IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2025.

Deep Reinforcement Learning-Based DRAM Equalizer Parameter Optimization Using Latent Representations

Under review at IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems

A deep reinforcement learning method that optimizes DRAM equalizer parameters directly in a learned latent representation space.

M. Usama and D. E. Chang, "Deep Reinforcement Learning-Based DRAM Equalizer Parameter Optimization Using Latent Representations," IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2025.

Patents

Method and Apparatus for Monitoring Health Status of Patient Based on Reliability Using Multiple Biosignals

Korea Patent Application 10-2025-0213058, filed December 29, 2025

A patent application for a method and apparatus that monitors patient health status by weighting multiple biosignals according to their reliability.

M. Usama and Y. R. Jung, "Method and Apparatus for Monitoring Health Status of Patient Based on Reliability Using Multiple Biosignals," Korea Patent Application 10-2025-0213058, filed Dec. 29, 2025. Applicant: BrainDeck Co., Ltd.

Reinforcement Learning-based Equalizer Parameter Optimization Method and System

Korea Patent Application 10-2025-0119661, filed August 26, 2025

A patent application for a reinforcement-learning-based method and system for optimizing DRAM equalizer parameters.

M. Usama and D. E. Chang, "Reinforcement Learning-based Equalizer Parameter Optimization Method and System," Korea Patent Application 10-2025-0119661, filed Aug. 26, 2025. Applicants: Samsung Electronics Co., Ltd. and KAIST.