TildAlice – Tech Blog for Developers
Python, AI, Deep Learning, Data Analysis, Computer Vision, Reinforcement Learning โ tips and tutorials
Latest Posts
-
FlashAttention ๋ฆฌ๋ทฐ: Transformer ์ดํ ์ 2~4๋ฐฐ ๊ฐ์
FlashAttention cuts Transformer memory from O(Nยฒ) to O(N) while running 2-4ร faster. Here's the IO-aware tiling trick explained.
-
RetNet ๋ ผ๋ฌธ ๋ฆฌ๋ทฐ: Transformer๋ฅผ ๋์ฒดํ ๋ณ๋ ฌ ํ์ต+์์ฐจ ์ถ๋ก ์ํคํ ์ฒ ์์ ๋ถ์
RetNet achieves 8.4ร faster inference than Transformer with 70% less memory. The parallel training + sequential inference proof inside.
-
[๋ ผ๋ฌธ๋ฆฌ๋ทฐ] Mamba: Selective State Space Model๋ก Transformer์ ํ๊ณ๋ฅผ ๋ํํ๋ค
Mamba beats Transformer with linear-time complexity using Selective State Space Models. See the selection mechanism that makes it work.
-
ConvNeXt vs Swin Transformer: ํ์ด๋ธ๋ฆฌ๋ ์ํคํ ์ฒ ์ค์ ์ฑ๋ฅ ๋น๊ต์ ์ต์ ์ ํ ์ ๋ต
ConvNeXt vs Swin Transformer on ImageNet, COCO, ADE20K: real accuracy and speed numbers to pick the right architecture for your project.
-
์ด๋ถ ํ์(Binary Search) ์๋ฒฝ ๊ฐ์ด๋ – ์๊ณ ๋ฆฌ์ฆ ์๋ฆฌ๋ถํฐ ์ฝ๋ฉํ ์คํธ ์์ฉ๊น์ง
Master binary search from O(log n) basics to parametric search. Includes lower/upper bound implementation and coding interview patterns.
-
RAG ํ์ดํ๋ผ์ธ ์ต์ ํ ์์ ๊ฐ์ด๋: Naive RAG๋ถํฐ Agentic RAG๊น์ง
Optimize RAG pipelines from naive to agentic: chunking strategies, hybrid search, and reranking code that actually improves retrieval.
-
XAI๋ก ๋ฒ ์ด๋ง ๊ณ ์ฅ ์์ธ ๋ถ์ํ๊ธฐ: SHAP๊ณผ Grad-CAM์ผ๋ก CNN ์ง๋จ ๋ชจ๋ธ ํด์
Use SHAP and Grad-CAM to prove your CNN bearing diagnosis model is looking at fault frequencies, not noise. Validation workflow included.
-
DeepSeek-V3 ๋ ผ๋ฌธ ๋ฆฌ๋ทฐ: MoE ๊ธฐ๋ฐ ์คํ์์ค LLM์ด GPT-4o๋ฅผ ๋์ด์ ๋น๊ฒฐ
DeepSeek-V3 matches GPT-4o with 671B parameters but activates only 37B. The MoE, MLA, and load balancing techniques that made it possible.