딥러닝 – TildAlice

Transformer Review: Attention Is All You Need 완전 분석

2026년 01월 04일

Transformer self-attention math decoded: positional encoding, multi-head mechanics, and why it beat RNN/LSTM. Full equations + ablations.

2026년 01월 03일

DDPM diffusion models explained: forward/reverse process math, simplified loss function derivation, and U-Net architecture that changed AI.

2025년 12월 17일

FlashAttention-3 cuts Transformer memory from O(n²) to O(n) on H100 GPUs. See PyTorch benchmarks proving 3-5x speedup on 32k context windows.

2025년 12월 10일

KAN uses 100× fewer parameters than MLP for function approximation. Kolmogorov-Arnold Networks: learnable activation functions on edges explained.

2025년 11월 28일

Fine-tune OpenAI Whisper for Korean speech: boost accuracy from 82% to 94% with custom datasets. Includes training code and benchmarks.

2025년 11월 17일

LoRA uses 60GB GPU, QLoRA needs 16GB, DoRA hits 12GB — same model quality. Here's the memory breakdown and when to pick each method.

2025년 11월 14일

YOLOv8 vs v11: real-world FPS and mAP benchmarks on edge devices. See which version actually wins for your hardware.

2025년 11월 07일

ViT or ConvNeXt for image classification in 2026? Speed vs accuracy tradeoffs tested on ImageNet with deployment recommendations.