Skip to content
@iq-ai-lab

IQ AI Lab

A math-first AI lab. We prove it, implement it from scratch, and document the insight.

๐Ÿ” IQ AI Lab

์ˆ˜ํ•™์  ์ฆ๋ช…์œผ๋กœ AI์˜ ๋ณธ์งˆ์„ ํƒ๊ตฌํ•˜๋Š” ๋”ฅ๋‹ค์ด๋ธŒ ์—ฐ๊ตฌ์†Œ


IQ AI Lab


"Use โ†’ Understand โ†’ Prove โ†’ Implement โ†’ Repeat"

ํ‘œ๋ฉด์ ์ธ ์‚ฌ์šฉ๋ฒ•์ด ์•„๋‹Œ,
์™œ ์ด ์ˆ˜์‹์ธ๊ฐ€ โ€” ๊ฐ€์ •๋ถ€ํ„ฐ ์ฆ๋ช…๊นŒ์ง€ ๋ฅผ ์ง‘์š”ํ•˜๊ฒŒ ํŒŒ๊ณ ๋“ญ๋‹ˆ๋‹ค.


๐Ÿ—บ๏ธ Learning Roadmap

graph TD
    M0["๐Ÿงฎ Layer 0\nMathematics"]
    M1["๐Ÿ“ Layer 1\nML Theory"]
    M2["๐Ÿง  Layer 2\nNeural Networks"]
    M3["๐Ÿ—๏ธ Layer 3\nArchitectures"]
    M4A["๐Ÿค– Layer 4\nReinforcement Learning"]
    M4B["๐Ÿ—ฃ๏ธ Layer 4\nLLM"]
    M4C["๐Ÿ‘๏ธ Layer 4\nComputer Vision"]
    M4D["๐Ÿ“ Layer 4\nNLP"]
    M5["โš™๏ธ Layer 5\nSystems"]

    M0 --> M1
    M1 --> M2
    M2 --> M3
    M3 --> M4A
    M3 --> M4B
    M3 --> M4C
    M3 --> M4D
    M4A --> M5
    M4B --> M5
    M4C --> M5
    M4D --> M5

    style M0 fill:#e3f2fd,stroke:#1565c0,stroke-width:2px
    style M1 fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px
    style M2 fill:#fff3e0,stroke:#ef6c00,stroke-width:2px
    style M3 fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px
    style M4A fill:#fce4ec,stroke:#c62828,stroke-width:2px
    style M4B fill:#fce4ec,stroke:#c62828,stroke-width:2px
    style M4C fill:#fce4ec,stroke:#c62828,stroke-width:2px
    style M4D fill:#fce4ec,stroke:#c62828,stroke-width:2px
    style M5 fill:#fff8e1,stroke:#fbc02d,stroke-width:2px
Loading

๐Ÿ“š Projects & Studies

๐Ÿงฎ Layer 0 โ€” Mathematics

AI์˜ ๋ชจ๋“  ์ˆ˜์‹์ด ์—ฌ๊ธฐ์„œ ์ถœ๋ฐœํ•ฉ๋‹ˆ๋‹ค

๐Ÿ“Œ Title ๐Ÿ“ Core Topics ๐Ÿ”— Link
Linear Algebra Deep Dive ๋ฒกํ„ฐ๊ณต๊ฐ„, ๊ณ ์œ ๊ฐ’ ๋ถ„ํ•ด, SVD, PCA ์ฆ๋ช…, ํ…์„œ ๐Ÿ“‚ Repository
Probability Theory Deep Dive ์ธก๋„ ๊ธฐ๋ฐ˜ ํ™•๋ฅ ๋ก , ์กฐ๊ฑด๋ถ€ ๊ธฐ๋Œ“๊ฐ’, ์ˆ˜๋ ด 4์ข…, ๋งˆํŒ…๊ฒŒ์ผ ๐Ÿ“‚ Repository
Mathematical Statistics Deep Dive MLE/MAP ์œ ๋„, ๊ฐ€์„ค๊ฒ€์ •, ์ ๊ทผ์ด๋ก , ๋ฒ ์ด์ฆˆ ์ถ”๋ก  ๐Ÿ“‚ Repository
Calculus & Optimization Deep Dive ํŽธ๋ฏธ๋ถ„, ์•ผ์ฝ”๋น„์•ˆ, ํ—ค์‹œ์•ˆ, ๋ผ๊ทธ๋ž‘์ฃผ ์Šน์ˆ˜๋ฒ• ๐Ÿ“‚ Repository
Convex Optimization Deep Dive ๋ณผ๋ก ์ง‘ํ•ฉ/ํ•จ์ˆ˜, KKT ์กฐ๊ฑด, ์Œ๋Œ€์ด๋ก , Proximal ๋ฐฉ๋ฒ• ๐Ÿ“‚ Repository
Information Theory Deep Dive Shannon Entropy, KL-Divergence, ์ƒํ˜ธ์ •๋ณด๋Ÿ‰, MDL ๐Ÿ“‚ Repository
Stochastic Processes Deep Dive ๋งˆ๋ฅด์ฝ”ํ”„ ์ฒด์ธ, ๋ธŒ๋ผ์šด ์šด๋™, ๋งˆํŒ…๊ฒŒ์ผ, MCMC ๐Ÿ“‚ Repository
Stochastic Differential Equations Deep Dive ์ดํ†  ์ ๋ถ„, ์ดํ†  ๊ณต์‹, SDE, Fokker-Planck ๋ฐฉ์ •์‹ ๐Ÿ“‚ Repository
Functional Analysis Deep Dive ํž๋ฒ ๋ฅดํŠธ ๊ณต๊ฐ„, ์ŠคํŽ™ํŠธ๋Ÿด ์ด๋ก , RKHS ๐Ÿ“‚ Repository
Information Geometry Deep Dive ํ†ต๊ณ„๋‹ค์–‘์ฒด, Fisher ์ •๋ณด ํ–‰๋ ฌ, Natural Gradient ๐Ÿ“‚ Repository

๐Ÿ“ Layer 1 โ€” ML Theory

๊ณ ์ „ ML์˜ ์ˆ˜ํ•™์  ํ† ๋Œ€์™€ ์ผ๋ฐ˜ํ™” ์ด๋ก 

๐Ÿ“Œ Title ๐Ÿ“ Core Topics ๐Ÿ”— Link
ML Fundamentals Deep Dive ์„ ํ˜•ํšŒ๊ท€(Normal Equation ์œ ๋„), ๊ฒฐ์ •ํŠธ๋ฆฌ, ์•™์ƒ๋ธ” ์ˆ˜๋ ด ์ฆ๋ช… ๐Ÿ“‚ Repository
Statistical Learning Theory Deep Dive PAC Learning, VC ์ฐจ์›, Rademacher ๋ณต์žก๋„, ๊ท ๋“ฑ์ˆ˜๋ ด ๐Ÿ“‚ Repository
Kernel Methods Deep Dive Mercer ์ •๋ฆฌ, RKHS, SVM ์Œ๋Œ€ ์œ ๋„, ๊ฐ€์šฐ์‹œ์•ˆ ํ”„๋กœ์„ธ์Šค ๐Ÿ“‚ Repository
Bayesian ML Deep Dive ๋ฒ ์ด์ง€์•ˆ ์ถ”๋ก , ๋ณ€๋ถ„ ๋ฒ ์ด์ฆˆ, ELBO ์œ ๋„, MCMC ๐Ÿ“‚ Repository
Graphical Models Deep Dive ๋ฒ ์ด์ง€์•ˆ ๋„คํŠธ์›Œํฌ, ๋งˆ๋ฅด์ฝ”ํ”„ ๋žœ๋ค ํ•„๋“œ, ๋ฏฟ์Œ ์ „ํŒŒ ๐Ÿ“‚ Repository

๐Ÿง  Layer 2 โ€” Neural Network Theory

์‹ ๊ฒฝ๋ง์˜ ์ˆ˜ํ•™์  ๋ณธ์งˆ โ€” ์—ญ์ „ํŒŒ๋ถ€ํ„ฐ ์ผ๋ฐ˜ํ™”๊นŒ์ง€

๐Ÿ“Œ Title ๐Ÿ“ Core Topics ๐Ÿ”— Link
Neural Network Theory Deep Dive ๋ฒ”์šฉ ๊ทผ์‚ฌ ์ •๋ฆฌ ์ฆ๋ช…, ์—ญ์ „ํŒŒ ์™„์ „ ์œ ๋„, Xavier/He ์ดˆ๊ธฐํ™” ์œ ๋„ ๐Ÿ“‚ Repository
Optimization Theory Deep Dive SGD ์ˆ˜๋ ด ๋ถ„์„, Adam ์ˆ˜๋ ด ๋ฐ˜๋ก€, Loss Landscape, LR Scheduler ๐Ÿ“‚ Repository
Generalization Theory Deep Dive Double Descent, Neural Tangent Kernel, Grokking, Lottery Ticket ๐Ÿ“‚ Repository
Regularization Theory Deep Dive L1/L2 ๋ฒ ์ด์ฆˆ ํ•ด์„, Dropout ์•™์ƒ๋ธ” ์ด๋ก , BatchNorm ์ฆ๋ช… ๐Ÿ“‚ Repository

๐Ÿ—๏ธ Layer 3 โ€” Architectures

ํ•ต์‹ฌ ์•„ํ‚คํ…์ฒ˜ ์ˆ˜์‹ ์™„์ „ ๋ถ„ํ•ด

๐Ÿ“Œ Title ๐Ÿ“ Core Topics ๐Ÿ”— Link
CNN Deep Dive Convolution ๋“ฑ๋ณ€์„ฑ, Receptive Field, ResNet ํ•ญ๋“ฑ ๊ทผ์‚ฌ ์ด๋ก  ๐Ÿ“‚ Repository
RNN & LSTM Deep Dive BPTT ์™„์ „ ์œ ๋„, Vanishing Gradient ์ŠคํŽ™ํŠธ๋Ÿด ๋ถ„์„, LSTM ๊ฒŒ์ดํŠธ ์ˆ˜์‹ ๐Ÿ“‚ Repository
Transformer Deep Dive Attention ์ˆ˜์‹ ์™„์ „ ๋ถ„ํ•ด, โˆšd ๊ทผ๊ฑฐ, Positional Encoding, Linear Attention ๐Ÿ“‚ Repository
Graph Neural Network Deep Dive ๊ทธ๋ž˜ํ”„ ๋ผํ”Œ๋ผ์‹œ์•ˆ ์œ ๋„, Message Passing, ํ‘œํ˜„๋ ฅ ํ•œ๊ณ„(WL ๊ฒ€์‚ฌ) ๐Ÿ“‚ Repository
Generative Model Deep Dive GAN minimax ์ˆ˜๋ ด, VAE ELBO ์™„์ „ ์œ ๋„, Normalizing Flow ๐Ÿ“‚ Repository

๐Ÿค– Layer 4-A โ€” Reinforcement Learning

๋ฒจ๋งŒ ๋ฐฉ์ •์‹๋ถ€ํ„ฐ PPO ๋‹จ์กฐ ๊ฐœ์„  ๋ณด์žฅ๊นŒ์ง€

๐Ÿ“Œ Title ๐Ÿ“ Core Topics ๐Ÿ”— Link
RL Foundations Deep Dive MDP ๊ณต๋ฆฌ์  ์ •์˜, ๋ฒจ๋งŒ ๋ฐฉ์ •์‹ ์™„์ „ ์œ ๋„, ๋™์  ํ”„๋กœ๊ทธ๋ž˜๋ฐ ์ˆ˜๋ ด ์ฆ๋ช… ๐Ÿ“‚ Repository
Model-Free RL Deep Dive Monte Carlo ์ˆ˜๋ ด ์กฐ๊ฑด, TD Learning, Q-Learning ์ˆ˜๋ ด ์ฆ๋ช…(Robbins-Monro) ๐Ÿ“‚ Repository
Deep RL Deep Dive DQN ๊ฒฝํ—˜ ์žฌ์ƒ ์ด๋ก , Double DQN ๊ณผ๋Œ€์ถ”์ • ํŽธํ–ฅ ์ฆ๋ช…, Rainbow ๐Ÿ“‚ Repository
Policy Gradient Deep Dive REINFORCE ์™„์ „ ์œ ๋„(๋กœ๊ทธ ๋ฏธ๋ถ„ ํŠธ๋ฆญ), ์ •์ฑ… ๊ธฐ์šธ๊ธฐ ์ •๋ฆฌ ์ฆ๋ช…, GAE ฮป-์ˆ˜์ต ์œ ๋„ ๐Ÿ“‚ Repository
Advanced RL Deep Dive TRPO ๋‹จ์กฐ ๊ฐœ์„  ์ด๋ก , PPO ํด๋ฆฌํ•‘ ๊ทผ๊ฑฐ, SAC ์—”ํŠธ๋กœํ”ผ ์ตœ๋Œ€ํ™”, TD3 ๐Ÿ“‚ Repository
RL Theory Deep Dive UCB ํ›„ํšŒ ๊ฒฝ๊ณ„ ์ฆ๋ช…, PAC-MDP, ํƒํ—˜-ํ™œ์šฉ ํŠธ๋ ˆ์ด๋“œ์˜คํ”„, ์ƒ˜ํ”Œ ๋ณต์žก๋„ ๐Ÿ“‚ Repository

๐Ÿ—ฃ๏ธ Layer 4-B โ€” Large Language Models

Scaling Law๋ถ€ํ„ฐ Alignment๊นŒ์ง€

๐Ÿ“Œ Title ๐Ÿ“ Core Topics ๐Ÿ”— Link
LLM Pretraining Deep Dive Scaling Law(Chinchilla ์ˆ˜์‹), ํ•™์Šต ์•ˆ์ •์„ฑ, ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ์ด๋ก  ๐Ÿ“‚ Repository
LLM Alignment Deep Dive RLHF(๋ณด์ƒ ๋ชจ๋ธ๋ง), DPO(Bradley-Terry ์œ ๋„), Constitutional AI ๐Ÿ“‚ Repository
LLM Efficiency Deep Dive LoRA(์ €๋žญํฌ ๋ถ„ํ•ด ์ˆ˜ํ•™), QLoRA ์–‘์žํ™” ์˜ค์ฐจ, MoE ์ˆ˜์‹, Speculative Decoding ๐Ÿ“‚ Repository
LLM Inference Deep Dive KV Cache, PagedAttention ๋ฉ”๋ชจ๋ฆฌ ๋ถ„์„, Continuous Batching, vLLM ๐Ÿ“‚ Repository

๐Ÿ‘๏ธ Layer 4-C โ€” Computer Vision

์ด๋ฏธ์ง€์˜ ์ˆ˜ํ•™์  ๊ตฌ์กฐ ํƒ๊ตฌ

๐Ÿ“Œ Title ๐Ÿ“ Core Topics ๐Ÿ”— Link
Vision Transformer Deep Dive ViT ํŒจ์น˜ ์ž„๋ฒ ๋”ฉ ์ˆ˜์‹, DINO ์ž๊ธฐ์ง€๋„, MAE ์ •๋ณด์ด๋ก ์  ํ•ด์„ ๐Ÿ“‚ Repository
Object Detection Deep Dive YOLO ์†์‹คํ•จ์ˆ˜ ์œ ๋„, DETR ํ—๊ฐ€๋ฆฌ์•ˆ ๋งค์นญ, Anchor-Free ๋ฐฉ๋ฒ•๋ก  ๐Ÿ“‚ Repository
Diffusion Model Deep Dive DDPM Forward/Reverse ์™„์ „ ์œ ๋„, Score Matching, DDIM ODE ํ•ด์„ ๐Ÿ“‚ Repository

๐Ÿ“ Layer 4-D โ€” NLP

์–ธ์–ด์˜ ์ˆ˜ํ•™์  ํ‘œํ˜„ ํƒ๊ตฌ

๐Ÿ“Œ Title ๐Ÿ“ Core Topics ๐Ÿ”— Link
NLP Foundations Deep Dive Word2Vec(Skip-gram ๋ชฉ์ ํ•จ์ˆ˜ ์œ ๋„), GloVe(ํ–‰๋ ฌ ๋ถ„ํ•ด ํ•ด์„), BPE ๐Ÿ“‚ Repository
Pretrained LM Deep Dive BERT MLM ์ •๋ณด์ด๋ก , GPT ์ž๊ธฐํšŒ๊ท€ ์ˆ˜์‹, ์ „์ดํ•™์Šต ์ด๋ก  ๐Ÿ“‚ Repository

โš™๏ธ Layer 5 โ€” Systems

์ˆ˜ํ•™์„ ์‹ค์ œ ์‹œ์Šคํ…œ์œผ๋กœ

๐Ÿ“Œ Title ๐Ÿ“ Core Topics ๐Ÿ”— Link
PyTorch Internals Deep Dive Autograd(์•ผ์ฝ”๋น„์•ˆ ๋ฒกํ„ฐ๊ณฑ), CUDA ์ปค๋„, ํ˜ผํ•ฉ ์ •๋ฐ€๋„ ์ˆ˜์น˜ ๋ถ„์„ ๐Ÿ“‚ Repository
Distributed Training Deep Dive Data/Model/Pipeline Parallelism, DDP AllReduce, ZeRO ๋ฉ”๋ชจ๋ฆฌ ๋ถ„์„ ๐Ÿ“‚ Repository
Efficient ML Deep Dive Pruning/Quantization/Distillation ์ˆ˜ํ•™, FlashAttention, vLLM ๐Ÿ“‚ Repository
MLOps Deep Dive ํ”ผ์ฒ˜ ์Šคํ† ์–ด, ๋ชจ๋ธ ๋ชจ๋‹ˆํ„ฐ๋ง, ๋ฐ์ดํ„ฐ ๋“œ๋ฆฌํ”„ํŠธ, A/B ํ…Œ์ŠคํŠธ ์ด๋ก  ๐Ÿ“‚ Repository

๐Ÿ’ก Layer 0 โ†’ Layer 5 ์ˆœ์„œ๋กœ ์„ ํ–‰ ์ง€์‹์ด ์Œ“์ž…๋‹ˆ๋‹ค. ๊ฐ ๋ ˆํฌ๋Š” ๋…๋ฆฝ์ ์œผ๋กœ๋„ ํ•™์Šต ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.


๐Ÿ› ๏ธ Study Method

graph LR
    A{{๐Ÿ” Explore}} -->|๋…ผ๋ฌธยท๋ ˆํผ๋Ÿฐ์Šค| B{{๐Ÿ“ Formulate}}
    B -->|์ˆ˜์‹ ์œ ๋„| C{{โœ๏ธ Prove}}
    C -->|๊ฐ€์ • ๊ฒ€์ฆ| D{{๐Ÿ’ป Implement}}
    D -->|NumPy ๊ฒ€์ฆ| E{{๐Ÿ“ Document}}
    E -.->|๋‹ค์Œ ์งˆ๋ฌธ| A

    style A fill:#e3f2fd,stroke:#1565c0,stroke-width:2px
    style B fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px
    style C fill:#fff3e0,stroke:#ef6c00,stroke-width:2px
    style D fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px
    style E fill:#fff8e1,stroke:#fbc02d,stroke-width:2px
Loading
Step Description
๐Ÿ” Explore ๋…ผ๋ฌธยท๊ณต์‹ ๋ ˆํผ๋Ÿฐ์Šค์—์„œ ํƒ๊ตฌํ•  ๊ฐœ๋… ์„ ์ •
๐Ÿ“ Formulate ์ˆ˜ํ•™์  ์„ ํ–‰ ์กฐ๊ฑด ์ •๋ฆฌ, ์ˆ˜์‹ ๊ตฌ์กฐ ํŒŒ์•…
โœ๏ธ Prove ๊ฐ€์ • โ†’ ๋ณด์กฐ์ •๋ฆฌ โ†’ ์ •๋ฆฌ, ํ•œ ์ค„์”ฉ ์™„์ „ ์œ ๋„
๐Ÿ’ป Implement ํ”„๋ ˆ์ž„์›Œํฌ ์—†์ด NumPy๋กœ ์ง์ ‘ ๊ตฌํ˜„ยท๊ฒ€์ฆ
๐Ÿ“ Document ์ฆ๋ช…ยท์ง๊ด€ยทํ•œ๊ณ„๋ฅผ ๋‚˜๋งŒ์˜ ์–ธ์–ด๋กœ ์ฒด๊ณ„์  ์ •๋ฆฌ

๐Ÿ’ก Philosophy

"์ˆ˜์‹์„ ์“ธ ์ค„ ์•„๋Š” ๊ฒƒ๊ณผ, ์ˆ˜์‹์ด ์™œ ๊ทธ๋ ‡๊ฒŒ ์ƒ๊ฒผ๋Š”์ง€ ์•„๋Š” ๊ฒƒ์€ ๋‹ค๋ฅด๋‹ค."

Why Math-First Deep Dive?

  • โœ๏ธ ์ฆ๋ช… ๊ฐ€๋Šฅํ•œ ์ดํ•ด - ๋ธ”๋ž™๋ฐ•์Šค ์—†์ด, ๊ฐ€์ •๋ถ€ํ„ฐ ๊ฒฐ๋ก ๊นŒ์ง€ ์ง์ ‘ ์œ ๋„ํ•ฉ๋‹ˆ๋‹ค
  • ๐Ÿ” ํ•œ๊ณ„ ์‹๋ณ„ - ์ˆ˜์‹์˜ ๊ฐ€์ •์ด ๋ฌด๋„ˆ์ง€๋Š” ์กฐ๊ฑด์„ ์ •ํ™•ํžˆ ์••๋‹ˆ๋‹ค
  • ๐Ÿ”— ๊ฐœ๋… ์—ฐ๊ฒฐ - ์„œ๋กœ ๋‹ค๋ฅธ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ๊ฐ™์€ ์ˆ˜ํ•™์  ๋ฟŒ๋ฆฌ์—์„œ ๋‚˜์˜ด์„ ๋ฐœ๊ฒฌํ•ฉ๋‹ˆ๋‹ค
  • ๐Ÿ’ป ๊ตฌํ˜„ ์‹ ๋ขฐ - ์ˆ˜์‹์„ ์ฝ”๋“œ๋กœ ์ง์ ‘ ์˜ฎ๊ธฐ๋ฉฐ ์ด๋ก ๊ณผ ์‹ค์ œ๋ฅผ ์—ฐ๊ฒฐํ•ฉ๋‹ˆ๋‹ค

๐Ÿ”— Related

IQ Dev Lab

AI ์‹œ์Šคํ…œ ๊ตฌํ˜„์˜ ๊ธฐ๋ฐ˜์ด ๋˜๋Š” ๊ฐœ๋ฐœ ๊ธฐ์ˆ ์€ IQ Dev Lab์—์„œ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค


โญ๏ธ ๋„์›€์ด ๋˜์…จ๋‹ค๋ฉด Star๋ฅผ ๋ˆŒ๋Ÿฌ์ฃผ์„ธ์š”!

Popular repositories Loading

  1. .github .github Public

  2. iq-ai-lab.github.io iq-ai-lab.github.io Public

    HTML

Repositories

Showing 2 of 2 repositories

Top languages

Loadingโ€ฆ

Most used topics

Loadingโ€ฆ