Tin Sum Cheng

· Developing generalization theory for kernel methods, with empirical validations on large-scale benchmarks · Pre-training and post-training of foundation models across Transformer, Mamba, and VLM architectures; designed evaluation pipelines and ablation studies · Analyzing training dynamics of optimizers (SGD, Adam, AdamW) through the lens of Neural Collapse — with accepted paper at ICLR 2026 · Cross-disciplinary research on molecular feature representations using kernel methods

· Optimized MatMul operators for next-generation NPU hardware · Resolved a 4-month-pending accuracy issue by implementing Triton Ascend kernels for Mamba · Implemented Triton Ascend kernels for Batch MatMul (BMM) · Built and shipped distributed CPU training workflows for simulation optimization · Curated training and evaluation datasets; delivered end-to-end learning solutions independently

ML/AI, Deep Learning Theory, Foundation Models, Optimization, AI for Science

Discrete Mathematics and Topology