Few-Shot Learning for Bird Species Classification
Master’s thesis applying few-shot learning to fine-grained bird species classification using meta-learning approaches.
Links
No public links available yet.
Problem
Bird species classification with scarce labels requires methods that generalize from few examples.
Solution
Implemented few-shot setups (N-way K-shot) with Prototypical/Matching/Relation Networks, episodic training, data augmentation, and analysis (confusion, t-SNE).
Impact
- Baseline comparison across FSL paradigms
- Reproducible pipeline for ecological datasets
- WIP: adding ablations and cross-domain tests
Tech Stack
Python, PyTorch/Lightning, torchvision, scikit-learn, NumPy/Pandas, Matplotlib, Weights & Biases (optional).
Related projects
Real-time emotion detection pipeline with CNN/DNN models using Python, Flask, and React.
Full-stack AI resume analyzer that scores ATS compatibility, detects skill gaps, matches keywords semantically, and rewrites experience bullets using a local NLP pipeline and cloud LLM.
AI-powered VC pitch simulator that stress-tests startup founders with a dual-role Gemini 2.5 Flash engine — acting as a hostile investor then switching to structured coach — across 5 investor personas and stage-adaptive difficulty levels.