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.

Posts by Collection

portfolio

Personalized Agentic RAG System

Adaptive retrieval system with autonomous agents, iterative self-correction, and user memory over domain-specific knowledge bases (LangGraph, LangChain, FAISS).

Language-Guided Navigation Agent

A 0.5B-parameter language model that learns to follow natural-language navigation instructions and answer questions about what it finds - built with an RLHF-style recipe in miniature: LoRA SFT on BFS expert demonstrations, then REINFORCE with a KL penalty against a frozen reference.

RL Research Portfolio

Self-directed reinforcement-learning research: DQN reproductions from scratch, an end-to-end RLHF pipeline for a 0.5B language model, and a learned world model (MLP ensemble) with CEM planning that reaches 100% success on a PyBullet robot arm and stays robust where PPO fails.

publications

Class-Adaptive Focal Sharpness-Aware Minimization for Long-Tailed Visual Recognition

Published in Manuscript in preparation (target: CVPR), 2026

CA-Focal-SAM redistributes sharpness-aware minimization at the optimizer level: a frequency-conditioned GammaAdaptor assigns per-class focal parameters, improving tail-class accuracy by 6.03% on CIFAR-10-LT (IR=200) with no architectural change and zero inference overhead.

Recommended citation: Appau, Gideon Kofi Amo, and Wang, Pengkun. (2026). "Class-Adaptive Focal Sharpness-Aware Minimization for Long-Tailed Visual Recognition." Manuscript in preparation.
Download Paper

Sharpness-Feedback Classwise Allocation for Long-Tailed Visual Recognition

Published in Manuscript in preparation (target: CVPR), 2026

SFCA combines a static log-scarcity prior with an online per-class SAM ascent-gap signal (robustly normalized EMA plus a bounded sigmoid mapping) to allocate per-class sharpness during training. On CIFAR-100-LT (IR=200, Logit Adjustment) it raises tail accuracy from 22.45% to 24.49% with zero inference overhead.

Recommended citation: Appau, Gideon Kofi Amo, and Wang, Pengkun. (2026). "Sharpness-Feedback Classwise Allocation for Long-Tailed Visual Recognition." Manuscript in preparation.
Download Paper

talks

Homomorphic Encryption Deployment on FPGA

Published:

Presented the design and deployment of a homomorphic-encryption workload on FPGA for the USTC engineering practice module, covering the resource/latency trade-offs of encrypted computation on reconfigurable hardware.

CA-Focal-SAM for Long-Tailed Visual Recognition

Published:

Presented the CA-Focal-SAM framework and its SFCA extension to the USTC research group: how frequency-conditioned per-class SAM weights and online sharpness feedback redistribute overall vs. tail performance on long-tailed benchmarks, with the multi-seed evidence and mechanism analyses.

teaching

Peer mentoring for junior master’s students

Mentorship, USTC, Software Engineering program, 2024

Served as a peer mentor for junior master’s students in the USTC Software Engineering program (2024 - 2025): guided PyTorch debugging and SAM optimizer implementations, and explained gradient mechanics and loss-landscape visualization to labmates.