Chang
Li.

Multi-label deep learning classifier for apple leaf diseases on the Plant Pathology 2021 dataset. Tackled class imbalance, multi-symptom co-occurrence, and uncontrolled field conditions.

• Built ResNet50 backbone (sigmoid + BCEWithLogitsLoss) on 18K orchard images with binary label encoding for 6 disease categories
• Systematic augmentation sweep (none → light → medium → heavy) identified medium augmentation as optimal — lifted Test F1 from 0.8808 → 0.9167
• Hyperparameter tuning across LR (5e-5–5e-4), batch size (32–128), and optimizers (Adam/AdamW/SGD) — Adam + LR 5e-4 + BS 32 won
• Per-class FN/FP + confusion-pair analysis identified "complex" (F1=0.669) as the primary generalization bottleneck
• 87.7% of 2,795 test samples predicted perfectly; 0% completely wrong

Addressed performance degradation under covariate shift by implementing Dynamic TTT — enabling a trained model to self-adapt per test sample without full retraining or ground-truth labels.

• Implemented TTT in PyTorch with self-supervised per-sample gradient updates targeting OOD blur and cartoon-style covariates
• Improved OOD accuracy from 51.17% → 80.83% (blur) and 45.67% → 71.50% (cartoon) — gains of ~29pp and ~26pp respectively
• Extended model with auxiliary transformation heads (rotation, blur, brightness) to guide structured feature realignment under covariate shift
• Automated large-scale synthetic covariate generation via Stable Diffusion 2 (Hugging Face) — benchmarked on Colab A100
• Designed full adaptive inference pipeline: generation → mini-training set construction → per-sample adaptation → evaluation

Built a DQN-powered autonomous agent to play Battle City Tanks — a stochastic, dynamic combat environment requiring coordinated attack-defense decision-making.

• Designed state vector encoding player/enemy positions and health; normalized for stable training across dynamic game states
• Implemented DQN with experience replay, epsilon-greedy exploration (ε decay), and target network stabilization (γ=0.95)
• Engineered multi-condition reward function: game-over penalties, enemy-kill bonuses, proximity rewards for strategic behavior
• Hybrid rule-based fallback (shooting heuristics + nearest-enemy targeting) supplements RL policy during early exploration
• Achieved consistent single-enemy victories; diagnosed state-space expansion and reward sparsity as multi-enemy scaling bottleneck