DOMAINS
知識領域
数学の基礎から実装まで、相互に依存しあう 8 つの領域が一本の完全な連鎖をなしています。
全 48 概念知識領域 · 8
数学と統計の基礎
ベクトル・行列・確率・微分 —— AI の言語
6 concepts
この領域が答える問い
- Why can data be written as a matrix?
- Which direction does the gradient actually point?
- Why can softmax be used as a probability?
機械学習
ルールを書き込むのではなく、データから規則を学ばせる
6 concepts
この領域が答える問い
- What separates “memorising” from “learning”?
- Why does a model that aces training fail in production?
- What can unlabelled data still be used for?
深層学習
特徴量設計を多層ネットワーク自身に任せる
6 concepts
この領域が答える問い
- What is a single neuron actually computing?
- How does backpropagation carry error backwards?
- Why are deeper networks harder to train?
自然言語処理と大規模言語モデル
分かち書きから Transformer、そして「話せる」大規模モデルへ
6 concepts
この領域が答える問い
- What exactly does a model see when it reads a word?
- Why does attention beat recurrence?
- Why does next-token prediction yield conversation?
コンピュータビジョン
画素から物体・シーン・三次元構造を読み取る
6 concepts
この領域が答える問い
- What does a convolution kernel actually learn?
- How do classification, detection and segmentation differ?
- How does a machine recover 3D from 2D images?
強化学習
試行錯誤と遅延報酬から一連の意思決定を学ぶ
6 concepts
この領域が答える問い
- Why does delayed reward make the problem hard?
- How do value functions and policy gradients divide the work?
- Why can RL improve large language models?
生成 AI
「これは何か」の判別から「どうあるべきか」の合成へ
6 concepts
この領域が答える問い
- What distribution is a generative model learning?
- Why is diffusion more stable than a GAN?
- How does a text prompt become an image?
AI エンジニアリング・安全性・倫理
動くモデルを、使えて信頼でき説明責任を果たせる製品へ
6 concepts
この領域が答える問い
- How are large models compressed onto consumer hardware?
- How do you evaluate when there is no answer key?
- Why is prompt injection so hard to eliminate?