Appau Gideon Kofi Amo

Appau Gideon Kofi Amo

M.S. student at USTC working on optimizer-level methods for long-tailed visual recognition and on world models for decision-making.

About me

I am Appau Gideon Kofi Amo, a master’s student in Software Engineering (AI/ML) at the University of Science and Technology of China (USTC), advised by Assoc. Res. Pengkun Wang, and a Chinese Government Scholarship (CSC) recipient. Previously, I earned a B.Sc. in Information Technology Education (First Class Honours, Top 10%) from the University of Education, Winneba, Ghana.

My research sits at the intersection of optimization and efficient learning. For my master’s thesis I study class-adaptive sharpness-aware minimization for long-tailed visual recognition: CA-Focal-SAM redistributes SAM’s per-class emphasis through a frequency-conditioned GammaAdaptor, and its SFCA extension adds online, geometry-driven feedback from SAM’s own ascent step, improving tail-class accuracy on CIFAR-100-LT (IR=200) from 22.45% to 24.49% with zero inference overhead. Alongside this, I maintain a self-directed reinforcement-learning portfolio, where a learned world model with CEM planning reaches 100% success on a robot-arm control task and stays robust under out-of-distribution physics shifts.

I am seeking funded PhD opportunities for Spring/Fall 2027 in efficient AI, long-tailed/computer vision, and world models & embodied AI. Here is my CV (PDF).

Research interests

  • Efficient and robust optimization: sharpness-aware minimization and long-tailed visual recognition
  • World models and model-based reinforcement learning for embodied AI
  • LLM agents, retrieval-augmented generation, and human feedback alignment

News

  • Aug 2026 - SFCA (Sharpness-Feedback Classwise Allocation) validated on CIFAR-100-C: strongest corrupted tail accuracy among compared checkpoints.
  • Jun 2026 - Completed mechanism-level study of CA-Focal-SAM on CIFAR-100-LT across imbalance ratios 50/100/200 with CE, Logit Adjustment, and LDAM outer objectives.
  • Mar 2026 - Added obstacle-avoidance CEM costs to the robot-arm world-model planner; collision rate drops from 57% to 23% where five PPO training configurations fail.
  • 2024 - Started M.S. at USTC as a Chinese Government Scholarship (CSC) recipient.