DeepFake detection with Deep Tree Network (DTN) architecture - 94.5% accuracy, 45 FPS real-time processing with zero-shot learning

11th place solution for the U-Tokyo Deep Learning Course MLP Competition (Top 0.8%). High-performance MLP implemented from scratch in NumPy, featuring AdamW, EMA, SWA, and MC Dropout.

15th place solution for the U-Tokyo Deep Learning Course Softmax Regression Competition (Top 0.9%). A highly optimized NumPy-only implementation featuring custom feature engineering (HOG/LBP), class-specific calibration, and distribution bias correction.

8th place solution for the U-Tokyo Deep Learning Course CIFAR-10 Competition (Top 0.6%). High-performance CNN pipeline featuring WideResNet-28-10, SAM, and advanced augmentation scheduling.

3rd place solution for the University of Tokyo Deep Learning Course Competition. VAE with U-Net-style skip connections and multi-seed checkpoint ensembling on FashionMNIST.