Sharpness-Feedback Classwise Allocation for Long-Tailed Visual Recognition
Appau, Gideon Kofi Amo, and Wang, Pengkun. (2026). "Sharpness-Feedback Classwise Allocation for Long-Tailed Visual Recognition." Manuscript in preparation.
CV
A current copy of my full CV is available here: Appau_Gideon_CV.pdf
Education
- M.S. in Software Engineering (AI/ML), University of Science and Technology of China, 2024 - 2027 (expected)
- Thesis: Class-Adaptive Focal Sharpness-Aware Minimization for Long-Tailed Visual Recognition. Advisor: Assoc. Res. Pengkun Wang.
- B.Sc. in Information Technology Education, University of Education, Winneba, Ghana, 2018 - 2022
- First Class Honours, Top 10%. Advisor: Dr. William Asiedu.
Work experience
- 2025 - Present: Graduate Researcher
- University of Science and Technology of China, Suzhou, China
- Proposed CA-Focal-SAM, a class-adaptive sharpness-aware minimization optimizer that improves tail-class accuracy by 6.03% on CIFAR-10-LT (IR=200) with no architectural change or inference overhead; designed GammaAdaptor (2 learnable scalars) mapping class frequency to per-class focal parameters, and a quadratic perturbation-radius curriculum verified by Hessian spectral analysis.
- Proposed SFCA (Sharpness-Feedback Classwise Allocation), combining a static log-scarcity prior with an online per-class SAM ascent-gap signal (robust EMA + bounded sigmoid mapping); raises CIFAR-100-LT (IR=200, Logit Adjustment) tail accuracy from 22.45% to 24.49%, and achieves the strongest corrupted tail accuracy on CIFAR-100-C.
- Multi-seed statistical protocol (3 runs, mean +/- std) on CIFAR-10/100-LT, ImageNet-LT, CUB-200-LT; 20+ qualitative analyses (t-SNE, 2D/3D loss landscapes, confusion matrices).
- Mar 2026 - Apr 2026: Machine Learning Intern
- Future Interns (remote)
- End-to-end time-series forecasting (6-month sales prediction); NLP ticket-classification pipeline with BERT/DistilBERT and sentiment-based priority routing; AI recruitment tool combining spaCy/NLTK resume parsing with LLM-based job matching.
- Nov 2022 - Aug 2023: IT Assistant
- Akenten Appiah Menkah University, Kumasi, Ghana
- Automated data workflows in Python (Pandas, NumPy), cutting manual processing time by 40%; optimized SQL databases with 5,000+ records, improving query performance by 30%.
- Aug 2021 - Mar 2022: Full-Stack Web Developer Intern
- Amidarr A-mole, Ghana
- Engineered 3 full-stack web applications (Flask, JavaScript, MySQL); integrated the PayStack payment API into an e-commerce platform.
Skills
- Languages: Python, JavaScript, SQL, Java, Bash
- ML/AI: PyTorch, scikit-learn, Transformers, LangChain, LangGraph, Ollama, AutoGen, FAISS, Chroma
- LLM/NLP: BERT, DistilBERT, spaCy, NLTK, RAG agents
- Data: Pandas, NumPy, Matplotlib, feature engineering, time-series forecasting
- Web & Tools: Flask, REST APIs, Gradio, MySQL, Git, Linux, Docker, LaTeX
Publications
Class-Adaptive Focal Sharpness-Aware Minimization for Long-Tailed Visual Recognition
Appau, Gideon Kofi Amo, and Wang, Pengkun. (2026). "Class-Adaptive Focal Sharpness-Aware Minimization for Long-Tailed Visual Recognition." Manuscript in preparation.
Talks
CA-Focal-SAM for Long-Tailed Visual Recognition
Seminar presentation at USTC Thesis Progress Seminar, Suzhou, Suzhou, China
Homomorphic Encryption Deployment on FPGA
Defense presentation at USTC Engineering Practice Defense, Suzhou, Suzhou, China
Teaching
Awards
- Chinese Government Scholarship (CSC), USTC, 2024
- First Class Honours, Top 10%, University of Education Winneba, 2022
- NVIDIA LLM RAG Agent Fundamentals, 2025
- DeepLearning.AI Advanced Learning Algorithms, 2025
- DeepLearning.AI Supervised Machine Learning: Regression & Classification, 2025