Structure

🗺️ TalentMe 技术基础与面试指南 Master Structure Map

⚠️ Auto-generated by scripts/generate-structure.mjs — do not edit manually.

目录树与主题清单来自磁盘遍历; aman-ai 源码映射与大纲维护于 structure-meta.json。

ADVERTISEMENT · 赞助推荐


📁 全局目录树结构 (Directory Tree Structure)

content/tech/
├── Foundations/ (技术数理与底层架构基石)
│   ├── ML/ (经典机器学习) — 6 个主题
│   ├── DL/ (深度学习基础) — 8 个主题
│   ├── LLM/ (大语言模型架构与对齐) — 10 个主题
│   ├── RL/ (强化学习与在线决策) — 5 个主题
│   ├── Multimodal/ (多模态与扩散生成) — 5 个主题
│   ├── AI_Infra/ (AI 基础设施、GPU 硬件与分布式并行) — 8 个主题
│   ├── AI_Engineering/ (AI 应用工程、Agent 智能体与 RAG) — 6 个主题
│   ├── System_Design/ (工业级系统设计) — 6 个主题
│   ├── Math/ (AI 数理基础) — 6 个主题
└── Interviews/ (面试岗位八股地图)
    ├── MLE/ (机器学习工程师) — 5 个主题
    ├── AIE/ (AI 系统/应用工程师) — 4 个主题
    ├── RS/ (算法研究员/科学家) — 4 个主题
    ├── DS/ (数据科学家) — 4 个主题

📖 模块:Foundations/ML

📄 clustering-and-knn

  • 对应 aman_ai 源码文件:clustering.md, knn.md, pca.md
  • 中文标题:无监督聚类与 KNN:K-Means++ 坐标下降、DBSCAN 密度聚类、GMM 期望最大化 (EM) 与 KD-Tree 极客指南
  • English Title: Unsupervised Clustering & KNN: K-Means++, DBSCAN, GMM-EM & KD-Tree Guide
  • 分类:foundations · 标签:clustering, k-means, dbscan, em-algorithm, gmm, knn, kd-tree
  • 章节核心大纲:
    1. K-Means (K-Means++ 初始化)、DBSCAN 密度聚类与 GMM 高斯混合模型
    1. KNN 距离度量与 KD-Tree / Ball-Tree 空间索引
    1. PCA 主成分分析、SVD 奇异值分解与方差最大化推导

📄 decision-trees-and-ensemble

  • 对应 aman_ai 源码文件:decision-trees.md, random-forest.md, xgboost.md, lightgbm.md
  • 中文标题:决策树与集成学习:CART、GBDT 负梯度拟合、XGBoost 二阶展开与 LightGBM 极客全解
  • English Title: Decision Trees & Ensemble Methods: CART, GBDT, XGBoost 2nd-Order & LightGBM Guide
  • 分类:foundations · 标签:decision-trees, gbdt, xgboost, lightgbm, ensemble-learning, random-forest
  • 章节核心大纲:
    1. CART 树 Gini 指数、ID3 信息增益比与剪枝策略
    1. Bagging 随机森林 (Random Forest) 与 Out-of-Bag (OOB) 评估
    1. Boosting 演进:GBDT 负梯度拟合、XGBoost 二阶泰勒展开与 LightGBM GOSS/EFB

📄 linear-and-logistic-regression

  • 对应 aman_ai 源码文件:linear-regression.md, logistic-regression.md
  • 中文标题:线性与逻辑回归:数理推导、Log-Odds、极大似然、VIF检测与 Bias-Variance 全景全解
  • English Title: Linear & Logistic Regression: Mathematical Derivations, Log-Odds, MLE, VIF & Bias-Variance Full Guide
  • 分类:foundations · 标签:linear-regression, logistic-regression, mle, vif, bias-variance, machine-learning
  • 章节核心大纲:
    1. 线性回归 OLS 闭式解公式推导与 Gauss-Markov 定理
    1. Ridge (L2) 与 Lasso (L1) 正则化几何与拉普拉斯/高斯先验等价性
    1. 逻辑回归 Sigmoid、Log-Odds 对数几率与 Cross-Entropy 损失求导

📄 ml-math-and-eval-metrics

  • 对应 aman_ai 源码文件:eval.md, sampling.md, calibrated-probabilities.md
  • 中文标题:机器学习评估指标与数据工程全景:分类/回归/排序(NDCG)、概率校准、不平衡采样与文本预处理极客指南
  • English Title: ML Evaluation Metrics & Data Engineering: Classification, Regression, Ranking (NDCG), Calibration & Preprocessing Guide
  • 分类:foundations · 标签:evaluation-metrics, ndcg, mrr, auc-roc, f-beta, smote, probability-calibration, preprocessing
  • 章节核心大纲:
    1. Precision, Recall, F1/F-beta, ROC-AUC 积分与 PR-AUC 曲线
    1. 不平衡数据采样:SMOTE 过采样与 Focal Loss 难易样本加权
    1. 概率校准 (Platt Scaling / Isotonic Regression) 与交叉验证

📄 probabilistic-models

  • 对应 aman_ai 源码文件:probabilistic-models.md, hmm.md, crf.md
  • 中文标题:概率图模型:朴素贝叶斯条件独立、HMM 维特比 (Viterbi) 动态规划与 CRF 解决标注偏置极客全解
  • English Title: Probabilistic Graphical Models: Naive Bayes, HMM Viterbi & Linear-Chain CRF Guide
  • 分类:foundations · 标签:naive-bayes, hmm, viterbi, crf, sequence-labeling, probabilistic-models
  • 章节核心大纲:
    1. 朴素贝叶斯分类器与拉普拉斯平滑 (Laplace Smoothing)
    1. HMM 隐马尔可夫模型三大问题 (评估、解码 Viterbi、学习 Baum-Welch)
    1. CRF 条件随机场势函数与全局正则化

📄 support-vector-machines

  • 对应 aman_ai 源码文件:svm.md, kernels.md
  • 中文标题:支持向量机 (SVM):最大间隔几何推导、对偶变换、KKT 条件与高斯 RBF 核技巧全解
  • English Title: Support Vector Machines (SVM): Max-Margin Geometry, Duality, KKT & RBF Kernel Guide
  • 分类:foundations · 标签:svm, duality, kkt-conditions, rbf-kernel, hinge-loss, machine-learning
  • 章节核心大纲:
    1. 凸二次规划问题、几何间隔与软间隔 C 惩罚项
    1. 拉格朗日对偶性、KKT 互补松弛性与支持向量判定
    1. Mercer 定理、核技巧 (Kernel Trick) 与 RBF / 多项式核函数

📖 模块:Foundations/DL

📄 activation-functions-and-gradients

  • 对应 aman_ai 源码文件:activation-functions.md, autograd.md, gradients.md
  • 中文标题:深度学习基础全景:激活函数族全演进(GELU/SwiGLU)、损失函数大一统 (CE/KL/Huber/InfoNCE/ArcFace) 与计算图反向传播极客指南
  • English Title: Deep Learning Foundations: Activations Evolution (GELU/SwiGLU), Loss Function Taxonomy (CE/KL/Huber/InfoNCE/ArcFace) & Autograd Backprop Guide
  • 分类:foundations · 标签:deep-learning, activation-functions, backpropagation, loss-functions, infonce, arcface, swiglu, kl-divergence, seo-optimized
  • 章节核心大纲:
    1. Autograd 前向/反向计算图与链式法则
    1. 激活函数演进:Sigmoid/Tanh $to$ ReLU $to$ LeakyReLU $to$ GELU $to$ SwiGLU
    1. 梯度消失 (Gradient Vanishing) 与梯度爆炸 (Explosion) 几何解释

📄 cnn-and-vit-architectures

  • 对应 aman_ai 源码文件:cnn.md, vit.md, resnet.md
  • 中文标题:视觉架构演进全景:2D 卷积、感受野迭代计算、Depthwise 深度可分离卷积、ResNet 恒等残差映射与 Vision Transformer (ViT) 极客指南
  • English Title: Vision Architectures Evolution: 2D Conv, Receptive Field Calculus, Depthwise Separable Conv, ResNet Identity Mapping & Vision Transformer (ViT) Guide
  • 分类:foundations · 标签:deep-learning, cnn, vision-transformer, vit, resnet, receptive-field, depthwise-separable-conv, seo-optimized
  • 章节核心大纲:
    1. Conv2D 感受野递推、Depthwise Separable Conv 深度可分离卷积
    1. ResNet 残差恒等映射 $mathbf{y} = mathcal{F}(mathbf{x}) + mathbf{x}$ 与梯度高速公路
    1. Vision Transformer (ViT) Patch Embedding、Position Embedding 与 Self-Attention

📄 debugging-and-dl-comp

  • 对应 aman_ai 源码文件:dl-debugging.md, grad-cam.md
  • 中文标题:深度学习调试与竞赛工程全景:4 步调试框架、单 Batch 过拟合验证、数值梯度检查、20大常见工程Bug、Grad-CAM 可解释性与架构归纳偏置选型指南
  • English Title: Deep Learning Debugging & Competition Engineering Taxonomy: 4-Step Debugging Framework, Single Batch Overfitting, Gradient Check & Grad-CAM Guide
  • 分类:foundations · 标签:deep-learning, model-debugging, sanity-check, gradient-checking, grad-cam, knowledge-distillation, inductive-bias, common-bugs, seo-optimized
  • 章节核心大纲:
    1. 深度学习模型调试 4 步框架 (Overfit Single Batch 验证)
    1. 梯度检查 (Numerical Gradient Check) 与数值不稳定性排查
    1. Grad-CAM 图像显著性热力图可视化与特征归因

📄 gan-and-generative-basics

  • 对应 aman_ai 源码文件:gan.md, wgan.md, vae.md
  • 中文标题:生成对抗网络 (GAN) 全景:Minimax 博弈、JS 散度缺陷、WGAN Wasserstein 距离推导、WGAN-GP 梯度惩罚与 Mode Collapse 极客指南
  • English Title: Generative Adversarial Networks (GAN) Taxonomy: Minimax Game, JS Divergence Flaw, WGAN Earth Mover Distance & WGAN-GP Guide
  • 分类:foundations · 标签:deep-learning, gan, wgan, wgan-gp, minimax-game, wasserstein-distance, js-divergence, mode-collapse, seo-optimized
  • 章节核心大纲:
    1. GAN Generator 与 Discriminator Minimax 零和博弈与最优判别器 $D^*(x)$
    1. JS 散度在不重叠分布下的梯度消失缺陷
    1. WGAN Wasserstein 距离推导与 WGAN-GP 1-Lipschitz 梯度惩罚

📄 gnn-and-graph-learning

  • 对应 aman_ai 源码文件:gnn.md, gcn.md, gat.md
  • 中文标题:图神经网络 (GNN) 全景:邻接矩阵、图拉普拉斯矩阵、消息传递机制 (MPNN)、GCN、GraphSAGE、GAT 与边特征建模极客指南
  • English Title: Graph Neural Networks (GNN) Taxonomy: Graph Laplacian, Message Passing (MPNN), GCN, GraphSAGE, GAT & Edge Feature Guide
  • 分类:foundations · 标签:deep-learning, gnn, gcn, graphsage, gat, message-passing, graph-laplacian, seo-optimized
  • 章节核心大纲:
    1. 图表示论、拉普拉斯矩阵 $L = D – A$ 与图谱卷积
    1. MPNN (Message Passing Neural Network) 聚合-更新范式
    1. GCN 重归一化拉普拉斯矩阵与 GAT 多头图注意力机制

📄 normalization-and-regularization

  • 对应 aman_ai 源码文件:layer-norm.md, rms-norm.md, dropout.md, regularization.md
  • 中文标题:归一化与正则化全景:BatchNorm、LayerNorm、RMSNorm、L0/L1/L2 权重衰减与 Inverted Dropout 极客指南
  • English Title: Normalization & Regularization Taxonomy: BatchNorm, LayerNorm, RMSNorm, L0/L1/L2 Weight Decay & Inverted Dropout Guide
  • 分类:foundations · 标签:deep-learning, batchnorm, layernorm, rmsnorm, l1-l2-regularization, dropout, feature-scaling, seo-optimized
  • 章节核心大纲:
    1. BatchNorm (Batch 维度) vs LayerNorm (Sequence 维度) 均值方差推导
    1. RMSNorm 均方根归一化简化与计算性能优势
    1. Pre-LN vs Post-LN 梯度流稳定性与 Inverted Dropout 机制

📄 optimizer-and-initialization

  • 对应 aman_ai 源码文件:optimizers.md, initialization.md, adamw.md
  • 中文标题:优化器与训练工程全景:SGD、Momentum、AdamW 解耦权重衰减、Xavier/Kaiming 初始化推导、梯度累积与重计算 (Checkpointing) 极客指南
  • English Title: Optimizers & Training Engineering Taxonomy: SGD, Momentum, AdamW Decoupled Weight Decay, Xavier/Kaiming Initialization & Gradient Checkpointing Guide
  • 分类:foundations · 标签:deep-learning, optimizer, adamw, xavier-initialization, kaiming-initialization, gradient-accumulation, gradient-checkpointing, hyperparameter-tuning, seo-optimized
  • 章节核心大纲:
    1. 优化器演进:SGD $to$ Momentum $to$ RMSprop $to$ Adam $to$ AdamW (解耦 Weight Decay)
    1. 权重初始化:Xavier/Glorot (Tanh) 与 Kaiming/He (ReLU) 前向/反向方差守恒证明
    1. 学习率 Scheduler (Cosine Annealing / Warmup) 策略

📄 rnn-lstm-and-mamba-ssm

  • 对应 aman_ai 源码文件:rnn.md, lstm.md, mamba.md, ssm.md
  • 中文标题:序列模型演进全景:RNN 随时间反向传播 (BPTT)、LSTM/GRU 门控机制、xLSTM 矩阵内存、HiPPO 矩阵与 Mamba 选择性状态空间模型 (S6) 极客指南
  • English Title: Sequence Models Evolution: RNN BPTT, LSTM/GRU Gating, xLSTM Matrix Memory, HiPPO Matrix & Mamba Selective SSM (S6) Guide
  • 分类:foundations · 标签:deep-learning, rnn, lstm, xlstm, mamba, state-space-model, ssm, hippo-matrix, rwkv, jamba, bptt, seo-optimized
  • 章节核心大纲:
    1. RNN BPTT 随时间反向传播与长距离依赖崩溃
    1. LSTM 遗忘门/输入门/输出门与加性 CTC 路径
    1. 状态空间模型 (SSM) 连续连续化与 Mamba (S6) 选择性硬件感知扫描

📖 模块:Foundations/LLM

📄 alignment-and-rlhf-dpo

  • 对应 aman_ai 源码文件:rlhf.md, dpo.md, grpo.md, ppo.md
  • 中文标题:大模型偏好对齐全景:RLHF 3 阶段、PPO 截断损失、DPO 隐式奖励代换、GRPO 与 PRM/ORPO 深度剖析
  • English Title: Preference Alignment: RLHF 3-Stage, PPO Clipped Loss, DPO Math Derivation, GRPO & PRM/ORPO
  • 分类:foundations · 标签:rlhf, ppo, dpo, grpo, orpo, kto, reward-model, gae, alignment
  • 章节核心大纲:
    1. RLHF 三阶段 (SFT $to$ Reward Model $to$ PPO) 架构
    1. DPO (Direct Preference Optimization) 偏好对齐闭式损失推导
    1. GRPO (Group Relative Policy Optimization) 组相对策略优化

📄 hallucination-and-factuality

  • 对应 aman_ai 源码文件:hallucination.md, factscore.md, ragas.md
  • 中文标题:大模型幻觉与真实性全景:内在/外在幻觉分类、FActScore 评估、RAGAS 框架与 RoPE 位置插值 (PI/NTK/YaRN) 扩展技术
  • English Title: LLM Hallucination & Factuality: Taxonomies, FActScore, RAGAS, SAFE & Context Extension (PI/NTK/YaRN)
  • 分类:foundations · 标签:hallucination, factuality, factscore, ragas, safe, position-interpolation, ntk-aware, yarn, needle-in-a-haystack
  • 章节核心大纲:
    1. Intrinsic / Extrinsic 幻觉诊断分类
    1. FActScore 原子事实拆解评估与 RAGAS 框架
    1. RoPE 位置插值 (PI/NTK/YaRN) 长上下文扩展

📄 llm-foundations-and-sota

  • 对应 aman_ai 源码文件:llm-sota.md, gpt4.md, llama.md, qwen.md
  • 中文标题:开源与商业 SOTA 大模型演进全景:从 BERT/GPT-4 到 LLaMA-3、Qwen-3、Gemma-4 与 Kimi-K2 架构对比
  • English Title: Open & Commercial SOTA LLM Evolution: From BERT/GPT-4 to LLaMA-3, Qwen-3, Gemma-4 & Kimi-K2
  • 分类:foundations · 标签:llm-sota, gpt-4, claude-4, gemini-2, llama-3, qwen-3, gemma-4, kimi-k2, tulu-3
  • 章节核心大纲:
    1. 开源与商业大模型演进全景
    1. LLaMA 3 / Qwen 2.5 / DeepSeek-V3 架构细节对比
    1. Dense vs MoE 架构 Scaling Law 曲线

📄 moe-architecture

  • 对应 aman_ai 源码文件:moe.md, deepseek-v3.md
  • 中文标题:MoE 混合专家模型与 DeepSeek MLA/MTP/mHC 架构解构:Top-k 门控、无辅助损失均衡、低秩潜注意力与 KAN 剖析
  • English Title: Mixture-of-Experts (MoE) & DeepSeek MLA/MTP/mHC Architecture: Top-k Routing, Aux-Loss-Free, KAN vs MLP
  • 分类:foundations · 标签:moe, deepseek-v3, deepseek-v4, mla, mtp, kan, gating-router, load-balancing
  • 章节核心大纲:
    1. MoE Top-k Gate 路由与 Aux Loss (负载均衡损失)
    1. DeepSeek-V3 多头潜在注意力 (MLA) 矩阵低秩投影
    1. MTP (Multi-Token Prediction) 多 Token 并行预测

📄 nlp-tasks-and-ner

  • 对应 aman_ai 源码文件:nlp-tasks.md, ner.md, cnns-for-text-classification.md
  • 中文标题:经典 NLP 任务全景:NER 命名实体识别、文本分类、seq2seq 翻译与文本蕴含 (NLI)
  • English Title: Classical NLP Tasks: NER, Text Classification, seq2seq Translation & NLI Entailment
  • 分类:foundations · 标签:nlp, ner, bilstm-crf, text-classification, translation, nli
  • 章节核心大纲:
    1. BiLSTM-CRF 序列标注与转移矩阵
    1. Viterbi 动态规划最优解码算子
    1. TextCNN 文本分类与 seq2seq 机器翻译

📄 parameter-efficient-fine-tuning

  • 对应 aman_ai 源码文件:peft.md, lora.md, qlora.md
  • 中文标题:高效微调 (PEFT) 架构全景:LoRA、QLoRA、DoRA、Prefix/Prompt Tuning、Adapters 与 MoRA/ReLoRA 深度剖析
  • English Title: Parameter-Efficient Fine-Tuning (PEFT): LoRA, QLoRA, DoRA, Prefix/Prompt Tuning, Adapters & MoRA/ReLoRA
  • 分类:foundations · 标签:peft, lora, qlora, dora, prefix-tuning, prompt-tuning, adapters, bitfit, mora, relora
  • 章节核心大纲:
    1. LoRA 参数更新 $Delta W = B cdot A$ 低秩矩阵分解推导
    1. QLoRA NF4 (NormalFloat 4) 量化与 Double Quantization 双重量化
    1. Prefix Tuning / P-Tuning v2 虚拟 Token 提示微调

📄 quantization-and-compression

  • 对应 aman_ai 源码文件:quantization.md, gptq.md, awq.md
  • 中文标题:大模型量化与模型压缩全景:INT8/INT4 映射、SmoothQuant 异常值平滑、GPTQ 二阶 Hessian 优化与 AWQ/知识蒸馏剖析
  • English Title: LLM Quantization & Model Compression: INT8/INT4 Mapping, SmoothQuant Outliers, GPTQ Hessian & AWQ/Distillation
  • 分类:foundations · 标签:quantization, int8, int4, smoothquant, gptq, awq, knowledge-distillation, model-compression
  • 章节核心大纲:
    1. 均匀量化 (Symmetric / Asymmetric) 标度 S 与零点 Z 计算
    1. GPTQ 基于黑塞矩阵 Inverse Hessian 的逐列量化与 AWQ 保护激活
    1. SmoothQuant 激活-权重平滑量化

📄 reasoning-and-cot

  • 对应 aman_ai 源码文件:reasoning.md, deepseek-r1.md, o1.md
  • 中文标题:推理大模型与慢思考全景:DeepSeek-R1 纯 RL 自进化、Aha Moment 顿悟、长 CoT 蒸馏与 OpenAI o1/o3 慢思考范式
  • English Title: Reasoning LLMs & Slow-Thinking: DeepSeek-R1 Pure RL, Aha Moment, Long CoT Distillation & OpenAI o1/o3
  • 分类:foundations · 标签:reasoning, deepseek-r1, openai-o1, chain-of-thought, slow-thinking, test-time-compute, grpo, cot-distillation
  • 章节核心大纲:
    1. CoT (Chain of Thought) 思维链与 Tree-of-Thought (ToT) 搜索
    1. DeepSeek-R1-Zero 纯 RL 自进化推理顿悟现象 (Aha Moment)
    1. 结果监督 (ORM) vs 过程监督 (PRM) 与小模型推理能力蒸馏

📄 tokenizer-and-sampling

  • 对应 aman_ai 源码文件:tokenizers.md, sampling.md, bpe.md
  • 中文标题:Tokenizer 分词器与 LLM 采样解码全景:BPE、WordPiece、SentencePiece、Temperature、Top-k/p、Min-p、Gumbel-Max、Penalty 与 Sequence Packing 打包优化
  • English Title: Tokenizer & Decoding Strategies: BPE, WordPiece, SentencePiece, Temperature, Top-k/p, Min-p, Gumbel-Max, Repetition Penalty & Sequence Packing
  • 分类:foundations · 标签:tokenizer, bpe, wordpiece, sentencepiece, temperature, top-p, top-k, min-p, gumbel-max, sequence-packing, flashattention-varlen
  • 章节核心大纲:
    1. Tokenizer 算法:BPE (Byte-Pair Encoding)、WordPiece 与 Unigram
    1. 文本解码采样:Temperature 缩放、Top-k、Top-p (Nucleus)
    1. Min-P 动态概率切片与 Repetition Penalty 重复惩罚

📄 transformer-architecture

  • 对应 aman_ai 源码文件:transformer.md, multi-head-attention.md, rope.md, flash-attention.md
  • 中文标题:Transformer 架构解构:Self-Attention、MHA/GQA/MQA、RoPE 与 FlashAttention 1/2/3 算子融合全景
  • English Title: Transformer Architecture Breakdown: Self-Attention, MHA/GQA/MQA, RoPE & FlashAttention 1/2/3 Operator Fusion
  • 分类:foundations · 标签:transformer, self-attention, flashattention, rope, gqa, mqa, encoder-decoder, bigbird
  • 章节核心大纲:
    1. Scaled Dot-Product Attention $frac{1}{sqrt{d_k}}$ 缩放因子推导与 MHA 多头注意力
    1. MQA (Multi-Query) 与 GQA (Grouped-Query) KV 共享机制
    1. RoPE 旋转位置编码复数旋转矩阵与 FlashAttention 1/2/3 Tiling 算子

📖 模块:Foundations/RL

📄 agentic-rl-and-reasoning-search

  • 对应 aman_ai 源码文件:agentic-rl.md, mcts.md, prm.md, rlvr.md
  • 中文标题:智能体 RL 与推理搜索全景:MCTS 蒙特卡洛树搜索、PRM 过程监督与 RLVR 可验证奖励
  • English Title: Agentic RL & Reasoning Search: MCTS, Process Supervision & RLVR
  • 分类:foundations · 标签:agentic-rl, mcts, prm, orm, rlvr, reasoning-search, deepseek-r1, trajectory-optimization
  • 章节核心大纲:
    1. Agentic RL 轨迹优化与 Hindsight 引导
    1. MCTS 蒙特卡洛树搜索与 AlphaGo / o1 / R1 演进
    1. PRM (过程监督) vs ORM (结果监督) 与 RLVR 可验证奖励

📄 bandits-and-online-decision

  • 对应 aman_ai 源码文件:mab.md, ucb.md, linucb.md, thompson-sampling.md
  • 中文标题:多臂老虎机与在线决策:MAB 探索与利用、LinUCB、Thompson Sampling 与 Contextual Bandits 落地
  • English Title: Multi-Armed Bandits & Online Decision: MAB, LinUCB & Contextual Bandits
  • 分类:foundations · 标签:bandits, mab, contextual-bandits, linucb, thompson-sampling, ucb, recsys
  • 章节核心大纲:
    1. MAB 基础:Exploration vs Exploitation 权衡
    1. UCB (Upper Confidence Bound) 与 Thompson Sampling 贝叶斯采样
    1. Contextual Bandits (LinUCB) 个性化推荐与搜索落地

📄 foundations-and-deep-rl

  • 对应 aman_ai 源码文件:rl.md, mdp.md, bellman.md, dqn.md, ppo.md
  • 中文标题:经典与深度强化学习全景:MDP 体系、Bellman 最优方程、DQN、Policy Gradient、PPO 与 SAC 原理解构
  • English Title: Foundations & Deep RL: MDP, Bellman Equations, DQN, Policy Gradient, PPO & SAC
  • 分类:foundations · 标签:rl-foundations, mdp, bellman-equation, dqn, policy-gradient, ppo, trpo, sac, actor-critic
  • 章节核心大纲:
    1. MDP 体系与 Bellman 期望/最优方程推导
    1. Value-Based: Q-Learning, DQN, Double DQN, Dueling DQN
    1. Policy-Based & Actor-Critic: REINFORCE, A2C, PPO, SAC, TD3

📄 model-based-rl-and-planning

  • 对应 aman_ai 源码文件:reinforcement-learning.md, world-models-jepa.md, robotics.md
  • 中文标题:基于模型的强化学习与规划:世界模型、Dyna、MPC、MuZero 与 Dreamer 全景
  • English Title: Model-Based RL & Planning: World Models, Dyna, MPC, MuZero & Dreamer
  • 分类:RL · 标签:model-based-rl, planning, world-models, mpc, muzero, dreamer, jepa, dyna
  • 章节核心大纲:
    1. 世界模型学习与 Dyna
    1. MPC 规划 (Random Shooting/CEM)
    1. MuZero/Dreamer 潜空间模型与复合误差

📄 offline-rl-and-imitation-learning

  • 对应 aman_ai 源码文件:deep-rl.md, reinforcement-learning.md, reinforcement-finetuning.md
  • 中文标题:离线强化学习与模仿学习:分布偏移、行为克隆、CQL、IQL 与 RLHF/DPO 全景全解
  • English Title: Offline RL & Imitation Learning: Distribution Shift, BC, CQL, IQL & the Road to RLHF/DPO
  • 分类:RL · 标签:offline-rl, imitation-learning, behavior-cloning, cql, iql, rlhf, dpo, off-policy-evaluation
  • 章节核心大纲:
    1. 分布偏移与 BC 局限
    1. CQL/IQL 保守离线学习
    1. Offline RL ↔ RLHF/DPO 联系

📖 模块:Foundations/Multimodal

📄 audio-and-speech-models

  • 对应 aman_ai 源码文件:audio.md, whisper.md
  • 中文标题:语音与音频处理全景:Whisper 弱监督架构、梅尔声谱图与 Audio-LLM 原理解构
  • English Title: Speech & Audio Processing: Whisper Architecture, Log-Mel Spectrogram & Audio-LLM
  • 分类:foundations · 标签:audio-processing, whisper, log-mel-spectrogram, speech-recognition, audio-llm, encodec
  • 章节核心大纲:
    1. 梅尔声谱图 (Log-Mel Spectrogram) 频域转换
    1. Whisper 弱监督 Encoder-Decoder 语音识别
    1. Audio-LLM 连续语音 Token 化与自回归生成

📄 clip-and-contrastive-learning

  • 对应 aman_ai 源码文件:clip.md, infonce.md
  • 中文标题:多模态对齐:CLIP 双塔对比学习、InfoNCE 损失、Zero-Shot 迁移与 SigLIP 原理解构
  • English Title: Multimodal Alignment: CLIP Dual-Tower Contrastive Learning, InfoNCE Loss & SigLIP
  • 分类:foundations · 标签:clip, contrastive-learning, infonce, siglip, multimodal-alignment, zero-shot
  • 章节核心大纲:
    1. CLIP 双塔 (Vision ViT + Text Encoder) 结构
    1. InfoNCE Loss 数理推导与 Temperature 缩放
    1. Zero-Shot 迁移分类与图文跨模态检索

📄 diffusion-models

  • 对应 aman_ai 源码文件:ddpm.md, ldm.md, sora.md
  • 中文标题:扩散模型全景:DDPM 数理推导、Latent Diffusion (LDM)、DiT 架构与 GPT-4o Native 生成
  • English Title: Diffusion Models: DDPM Derivation, Latent Diffusion, DiT & GPT-4o Native Generation
  • 分类:foundations · 标签:diffusion-models, ddpm, stable-diffusion, dit, latent-diffusion, sora
  • 章节核心大纲:
    1. DDPM 正向加噪 Markov 链与反向去噪 U-Net 预测器
    1. Latent Diffusion (LDM) 潜空间压缩与 Stable Diffusion
    1. Native 图像/视频生成 (GPT-4o, Sora)

📄 vision-language-models

  • 对应 aman_ai 源码文件:vlm.md, llava.md, janus.md
  • 中文标题:视觉语言大模型 VLM 全景:ViT、Projector 桥接层、LLaVA 阶段训练与 DeepSeek-Janus Pro 原理解构
  • English Title: Vision-Language Models (VLM): ViT, Projectors, LLaVA 2-Stage & DeepSeek-Janus Pro
  • 分类:foundations · 标签:vlm, llava, deepseek-janus, q-former, projector, multimodal-llm
  • 章节核心大纲:
    1. VLM 经典三件套 (ViT + Projector + LLM)
    1. LLaVA 两阶段预训练与指令微调
    1. DeepSeek-Janus Pro 解耦表征与统一生成

📄 world-models-jepa

  • 对应 aman_ai 源码文件:world-models.md, jepa.md
  • 中文标题:世界模型与 JEPA 全景:Yann LeCun 非生成式表征预测、I-JEPA / V-JEPA 与具身智能 (VLA) 落地
  • English Title: World Models & JEPA: Yann LeCun’s Non-Generative Prediction, I-JEPA/V-JEPA & Embodied AI (VLA)
  • 分类:foundations · 标签:world-models, jepa, i-jepa, v-jepa, yann-lecun, embodied-ai, robotics
  • 章节核心大纲:
    1. Yann LeCun JEPA (I-JEPA / V-JEPA) 非生成式表征预测
    1. 世界模型对物理世界动态的隐式模拟
    1. 具身智能 (Embodied AI) VLA 机器人操纵

📖 模块:Foundations/AI_Infra

📄 cluster-scheduling-and-ray

  • 对应 aman_ai 源码文件:infra/kubernetes.md, infra/docker.md, ml-runtimes.md
  • 中文标题:集群调度与 Ray:K8s 调度管线、Raylet 架构、分布式对象存储、弹性扩缩容与 GPU 调度全景
  • English Title: Cluster Scheduling & Ray: K8s Scheduling Pipeline, Raylet Architecture, Object Store, Autoscaling & GPU Scheduling Full Guide
  • 分类:AI_Infra · 标签:ray, cluster-scheduling, kubernetes, autoscaler, gpu-scheduling, distributed-training, fault-tolerance, actor
  • 章节核心大纲:
    1. 集群调度问题与 DRF 公平性
    1. Kubernetes 调度原理与 Ray 对比
    1. Ray 架构/弹性扩缩容/GPU 调度/故障恢复

📄 distributed-parallelism

  • 对应 aman_ai 源码文件:— (新增主题,待补充)
  • 中文标题:分布式并行训练全景:TP 张量并行、PP 流水线并行、DP 数据并行与 DeepSpeed ZeRO 1/2/3
  • English Title: Distributed Training Parallelism: TP, PP, DP & DeepSpeed ZeRO 1/2/3
  • 分类:foundations · 标签:distributed-training, tensor-parallelism, pipeline-parallelism, deepspeed-zero, megatron-lm, 4d-parallelism

📄 gpu-hardware-and-hbm

  • 对应 aman_ai 源码文件:gpu-architecture.md, matmul.md
  • 中文标题:GPU 硬件架构全景:SM 流处理器、Tensor Core 混合精度、HBM 带宽与 Roofline 模型
  • English Title: GPU Hardware Architecture: SM, Tensor Cores, HBM Bandwidth & Roofline Model
  • 分类:foundations · 标签:gpu-architecture, hbm, tensor-cores, roofline-model, flashattention, cuda, h100
  • 章节核心大纲:
    1. Streaming Multiprocessor (SM)、Shared Memory 与 HBM 显存物理结构
    1. Tensor Cores 混合精度 GEMM 矩阵乘法 (MMA/WGMMA 指令集)
    1. Roofline Model 算术强度 $I = text{FLOPs}/text{Bytes}$ 与 Memory-Bound 判定

📄 high-concurrency-ai-system

  • 对应 aman_ai 源码文件:— (新增主题,待补充)
  • 中文标题:高并发 AI 系统设计全景:SSE 流式打字机推送、语义缓存 (Semantic Cache) 与 ML Runtimes
  • English Title: High-Concurrency AI System Design: SSE Streaming, Semantic Cache & ML Runtimes
  • 分类:foundations · 标签:system-design, sse, semantic-cache, high-concurrency, triton, tensorrt-llm, llm-serving

📄 kv-cache-and-paged-attention

  • 对应 aman_ai 源码文件:speculative-decoding.md, kv-cache-management.md
  • 中文标题:KV Cache 显存管理全景:内存碎片拟合推导、vLLM PagedAttention 虚拟内存与 Prefix Caching
  • English Title: KV Cache Management: Exact Bounds Derivation, vLLM PagedAttention & Prefix Caching
  • 分类:foundations · 标签:kv-cache, paged-attention, vllm, memory-management, prefix-caching, llm-inference
  • 章节核心大纲:
    1. KV Cache 容量推导与内外内存碎片消除
    1. vLLM PagedAttention 逻辑页表到物理 Block 映射与 Copy-on-Write
    1. Prefix Caching (Radix Tree) 与 Chunked Prefill TTFT 优化

📄 mlops-and-testing

  • 对应 aman_ai 源码文件:— (新增主题,待补充)
  • 中文标题:MLOps 与在线测试全景:Data Drift 监控、PSI 指标、A/B 测试与 CUPED 方差降低
  • English Title: MLOps & Online Testing: Data Drift Monitoring, PSI Metric, A/B Testing & CUPED
  • 分类:foundations · 标签:mlops, llmops, data-drift, psi, ab-testing, cuped, model-monitoring

📄 security-and-privacy

  • 对应 aman_ai 源码文件:— (新增主题,待补充)
  • 中文标题:AI 安全与隐私全景:Prompt 注入攻击、Guardrails 防御、差分隐私与联邦学习
  • English Title: AI Safety & Privacy: Prompt Injection, Guardrails, Differential Privacy & Federated Learning
  • 分类:foundations · 标签:ai-safety, prompt-injection, guardrails, differential-privacy, federated-learning, pii-masking

📄 speculative-decoding

  • 对应 aman_ai 源码文件:— (新增主题,待补充)
  • 中文标题:猜想解码 (Speculative Decoding) 全景:Draft Model 小模型草稿、拒绝采样证明与端侧加速
  • English Title: Speculative Decoding: Draft Model Sampling, Rejection Sampling & On-Device Acceleration
  • 分类:foundations · 标签:speculative-decoding, draft-model, rejection-sampling, llm-acceleration, on-device-ai, medusa

📖 模块:Foundations/AI_Engineering

📄 agent-design-patterns

  • 对应 aman_ai 源码文件:agents.md, agentic-design-patterns.md, langgraph.md
  • 中文标题:Agent 设计模式全景:ReAct 循环、Reflexion 自我反思、Plan-and-Execute 与 LangGraph 图工程
  • English Title: Agent Design Patterns: ReAct Loop, Reflexion Self-Correction, Plan-and-Execute & Graph Engineering
  • 分类:foundations · 标签:agent, react, reflexion, plan-and-execute, langgraph, claude-code, agentic-patterns
  • 章节核心大纲:
    1. ReAct (Reasoning + Acting) 循环与 Prompt Parsing
    1. Plan-and-Execute / Reflexion 自自我反思圈
    1. LangGraph / AutoGPT 状态图工程与 Claude Code

📄 llm-as-a-judge

  • 对应 aman_ai 源码文件:— (新增主题,待补充)
  • 中文标题:LLM-as-a-Judge 自动化评估全景:Pointwise 与 Pairwise 范式、三大 Bias 消除与 Cohen’s Kappa 统计一致性
  • English Title: LLM-as-a-Judge Evaluation: Pointwise & Pairwise Paradigms, Bias Elimination & Cohen’s Kappa
  • 分类:foundations · 标签:llm-as-a-judge, evaluation, cohens-kappa, benchmarks, elo-rating, bias-mitigation

📄 naive-and-advanced-rag

  • 对应 aman_ai 源码文件:rag.md, advanced-rag.md, hyde.md
  • 中文标题:RAG 检索增强生成全景:从 Naive RAG 到 Advanced RAG、混合检索 (BM25 + Dense)、RRF 与 Cross-Encoder 重排序
  • English Title: RAG Pipeline: From Naive RAG to Advanced RAG Architecture, Hybrid Search, RRF & Cross-Encoder
  • 分类:foundations · 标签:rag, advanced-rag, bm25, hybrid-search, rrf, reranking, hyde
  • 章节核心大纲:
    1. Fixed-size / Parent-Document Chunking 策略
    1. BM25 稀疏 + Dense 稠密向量混合检索与 RRF 融合
    1. HyDE 假设性文档嵌入与 Cross-Encoder 重排序

📄 prompt-engineering-and-guardrails

  • 对应 aman_ai 源码文件:prompt-engineering.md, attacks.md, pii.md
  • 中文标题:Prompt 工程与安全护栏:Structured Outputs、Outlines 语法硬约束与 Llama Guard 防护
  • English Title: Prompt Engineering & Safety Guardrails: Outlines & Llama Guard
  • 分类:AI_Engineering · 标签:prompt-engineering, guardrails, structured-outputs, outlines, llama-guard
  • 章节核心大纲:
    1. System Prompt 设计原则与 Few-shot 示例
    1. Structured Outputs 语法硬约束 (Outlines & Logit Bias)
    1. Llama Guard 越狱防护、PII 脱敏与重试机制

📄 tool-use-and-function-calling

  • 对应 aman_ai 源码文件:toolformer.md, Toolformer.md
  • 中文标题:工具调用与 Function Calling 全景:Toolformer 自主插入、JSON Schema 规范与沙箱安全执行
  • English Title: Tool Use & Function Calling: Toolformer Self-Taught Calls, JSON Schema & Sandbox Execution
  • 分类:foundations · 标签:function-calling, tool-use, toolformer, json-schema, sandbox, code-interpreter
  • 章节核心大纲:
    1. Toolformer 自动 API 提示生成与 Loss 过滤
    1. OpenAI Function Calling 规范与 TaskMatrix
    1. 代码解释器与沙箱安全隔离执行

📄 vector-databases-and-hnsw

  • 对应 aman_ai 源码文件:vector-dbs.md, embeddings.md, ann-similarity-search.md
  • 中文标题:Vector DB 向量数据库全景:HNSW 图索引、IVF-PQ 乘积量化与 ANN 相似度检索原理解构
  • English Title: Vector Databases: HNSW Graph Indexing, IVF-PQ Quantization & ANN Similarity Search
  • 分类:foundations · 标签:vector-db, hnsw, ivf-pq, ann, embeddings, milvus, qdrant
  • 章节核心大纲:
    1. Embeddings 生成与 ANN 近似最近邻检索
    1. IVF 倒排网格、PQ 乘积量化与 HNSW 多层小世界图
    1. Milvus / Qdrant 生产级选型与标量过滤

📖 模块:Foundations/System_Design

📄 industry-case-studies-pinterest-netflix

  • 对应 aman_ai 源码文件:pinterest.md, netflix.md, uber.md
  • 中文标题:业界经典 System Case Studies:Pinterest 视觉搜索与 Netflix 推荐系统
  • English Title: Industry System Case Studies: Pinterest Visual Search & Netflix Recommendation
  • 分类:foundations · 标签:system-design, case-study, pinterest, netflix, pinsage, ann-embedding, recommendation, ab-testing
  • 章节核心大纲:
    1. Pinterest 画板图推荐系统与大规模分布式 PinSage 架构
    1. Netflix 全球 CDN 视频流式传输与个性化封面生成系统
    1. Uber 实时派单调度、H3 六边形地理索引与动态加价系统

📄 llm-rag-agent-system-design

  • 对应 aman_ai 源码文件:sys-design.md, LLMOps.md
  • 中文标题:大模型 RAG 与 Agent 生产级系统架构:多租户隔离、流式服务与高可用
  • English Title: Production LLM RAG & Agent System Design: Multi-Tenancy, SSE & High Availability
  • 分类:foundations · 标签:system-design, rag-system-design, agent-system-design, multi-tenancy, sse-streaming, semantic-cache, context-engineering, ragas-eval
  • 章节核心大纲:
    1. 企业级知识库 RAG 混合架构设计
    1. 多租户数据隔离与细粒度权限控制
    1. HTTP SSE 流式打字机推送、Semantic Cache 语义缓存与降级熔断

📄 multimodal-generative-system-design

  • 对应 aman_ai 源码文件:sys-design.md, sora.md
  • 中文标题:多模态生成系统架构设计:图像/视频生成服务、模型切片与 GPU 动态扩缩容
  • English Title: Multimodal Generative System Design: Image/Video Generation & GPU Scaling
  • 分类:foundations · 标签:system-design, multimodal-system-design, diffusion-serving, task-queue, gpu-scaling, text-to-image, inference-acceleration, content-safety
  • 章节核心大纲:
    1. Stable Diffusion / Sora 视频生成分布式推理架构
    1. 大模型切片装载与显存 Swap 动态调度
    1. 高并发异步队列 (Celery/RabbitMQ) 与 GPU 动态 Auto-Scaling

📄 recommendation-system-design

  • 对应 aman_ai 源码文件:sys-design.md, pinterest.md
  • 中文标题:推荐系统工业级架构设计:召回-精排-重排三阶段、双塔模型与离在线一致性 Feature Store
  • English Title: Industry Recommendation System Design: 3-Stage Pipeline, Two-Tower Models & Feature Store
  • 分类:foundations · 标签:system-design, recommendation-system, two-tower-model, mmoe, feature-store, ranking, deepfm, dcn
  • 章节核心大纲:
    1. 召回 (Retrieval) $to$ 粗排 $to$ 精排 (Heavy Ranking) $to$ 重排三阶段漏斗
    1. DSSM 双塔模型向量化召回与 ANN 预索引
    1. Feature Store 离在线特征一致性 (Kafka+Flink+Redis) 与 MMoE 多目标优化

📄 risk-control-and-fraud-detection

  • 对应 aman_ai 源码文件:sys-design.md, fraud.md
  • 中文标题:实时风控与欺诈检测系统架构:流批一体、图风控与实时特征工程
  • English Title: Real-time Risk Control & Fraud Detection System Design: Streaming & Graph Risk
  • 分类:foundations · 标签:system-design, risk-control, fraud-detection, flink, graph-risk, rule-engine, anomaly-detection, imbalanced-learning
  • 章节核心大纲:
    1. 实时风控引擎架构:低延迟规则引擎与模型评分
    1. 流批一体 (Flink/Kafka) 毫秒级窗口特征计算
    1. 图风控 (Graph Risk) 黑灰产团伙识别与关系网络

📄 search-and-ad-system-design

  • 对应 aman_ai 源码文件:sys-design.md, search.md
  • 中文标题:搜索与计算广告系统设计:Query 意图理解、分布式倒排索引、RTB 竞价与 pCTR 预估
  • English Title: Search & Advertising System Design: Query Understanding, Inverted Index, RTB & pCTR Prediction
  • 分类:foundations · 标签:system-design, search-engine, advertising-system, rtb, inverted-index, pctr, learning-to-rank, gsp-auction
  • 章节核心大纲:
    1. 搜索引擎架构:Query 扩展、意图识别与分布式倒排索引
    1. 计算广告 RTB (Real-Time Bidding) 实时竞价机制与 10ms 拍卖
    1. pCTR / pCVR 预估与 AUC 监控

📖 模块:Foundations/Math

📄 learning-paradigms-and-bias

  • 对应 aman_ai 源码文件:inductive-bias.md, double-descent.md, learning-paradigms.md, fundamentals.md, learning-strategy.md
  • 中文标题:泛化理论全景:归纳偏置 (Inductive Bias)、Double Descent 双重下降与 PAC 学习范式
  • English Title: Generalization Theory: Inductive Bias, Double Descent & PAC Learning Paradigms
  • 分类:foundations · 标签:math, inductive-bias, double-descent, pac-learning, generalization, learning-paradigms
  • 章节核心大纲:
    1. 归纳偏置 (Inductive Bias) 在各神经网络架构中的定义
    1. 经典拟合曲线 vs 深度学习 Double Descent 现象
    1. AI 核心基础概念 (Fundamentals) 与学习范式对比

📄 linear-algebra-for-ai

  • 对应 aman_ai 源码文件:fundamentals.md, primers/math 矩阵分解/Jacobian-Hessian sections
  • 中文标题:线性代数核心:向量空间、四大子空间、SVD/EVD、投影与最小二乘、Jacobian/Hessian 全景
  • English Title: Linear Algebra Core for AI: Vector Spaces, Four Subspaces, EVD/SVD, Projection & Least Squares, Jacobian/Hessian
  • 分类:Math · 标签:linear-algebra, svd, eigenvalue, four-subspaces, least-squares, projection, pseudoinverse, jacobian
  • 章节核心大纲:
    1. 向量空间/四大子空间与秩
    1. EVD/SVD 分解与几何意义
    1. 投影、最小二乘与 PCA/低秩近似

📄 optimization-and-matrix-calculus

  • 对应 aman_ai 源码文件:gradient-descent.md, gradient-descent-1.md, bayesian-optimization.md, matmul.md
  • 中文标题:凸优化与矩阵求导全景:拉格朗日乘子法、KKT 条件、SVD 奇异值分解与梯度几何收敛
  • English Title: Optimization & Matrix Calculus: Lagrange Multipliers, KKT Conditions, SVD & Convergence Geometry
  • 分类:foundations · 标签:math, optimization, kkt-conditions, svd, matrix-calculus, convexity
  • 章节核心大纲:
    1. 矩阵求导法则 (Jacobian & Hessian) 与二次型
    1. 拉格朗日乘子法与 KKT 4 大条件 (互补松弛性)
    1. SVD 奇异值分解、Eckart-Young 低秩近似与二阶牛顿法收敛

📄 probability-and-information-theory

  • 对应 aman_ai 源码文件:bayes-theorem.md
  • 中文标题:AI 数理基础全景:贝叶斯推断、香农信息熵、交叉熵与 KL 散度非对称证明
  • English Title: AI Math Foundations: Bayes Inference, Shannon Entropy, Cross-Entropy & KL Divergence
  • 分类:foundations · 标签:math, bayes-theorem, entropy, cross-entropy, kl-divergence, information-theory
  • 章节核心大纲:
    1. 先验、似然与后验概率:贝叶斯定理应用
    1. 香农信息量、信息熵 (Entropy) 与交叉熵
    1. KL 散度非对称证明与分布匹配

📄 sampling-and-monte-carlo

  • 对应 aman_ai 源码文件:the-bootstrap.md, data-sampling.md, token-sampling.md
  • 中文标题:采样与蒙特卡洛方法:逆变换采样、拒绝采样、重要性采样、MCMC 与 Bootstrap 全景
  • English Title: Sampling & Monte Carlo Methods: Inverse Transform, Rejection, Importance Sampling, MCMC & Bootstrap
  • 分类:Math · 标签:sampling, monte-carlo, mcmc, bootstrap, importance-sampling, rejection-sampling, variance-reduction, llm-decoding
  • 章节核心大纲:
    1. 逆变换/拒绝/重要性采样
    1. MCMC (MH/Gibbs) 与诊断
    1. Bootstrap/蒙特卡洛积分/降方差

📄 statistics-and-hypothesis-testing

  • 对应 aman_ai 源码文件:gaussiannormal-distribution.md, central-limit-theorem.md, students-t-distribution.md, confidence-intervals.md
  • 中文标题:统计推断与假设检验:分布族、极大似然、中心极限定理、p-value、置信区间与功效分析全景
  • English Title: Statistical Inference & Hypothesis Testing: Distribution Families, MLE, CLT, p-Values, Confidence Intervals & Power Analysis
  • 分类:Math · 标签:statistics, hypothesis-testing, clt, mle, confidence-intervals, p-value, power-analysis, bootstrap
  • 章节核心大纲:
    1. 常用分布族与 CLT
    1. 点估计 MLE/矩估计与性质
    1. 假设检验/p-value/多重比较/功效分析

📖 模块:Interviews/MLE

📄 mle-coding-and-algo-prep

  • 对应 aman_ai 源码文件:— (新增主题,待补充)
  • 中文标题:MLE 算法手写实战:零基础 Pure Numpy 手写 LR、K-Means、Self-Attention 与 NMS
  • English Title: MLE Coding & Algo Prep: Zero-to-One ML Operators in Pure Numpy
  • 分类:MLE · 标签:mle-coding, numpy, handwritten-ml, self-attention, nms

📄 mle-core-cheatsheet

  • 对应 aman_ai 源码文件:interview.md, ml-comp.md, data-filtering.md, data-sampling.md, data-split.md, tips.md, pinterest.md
  • 中文标题:MLE 机器学习工程师核心地图:高频八股、竞赛经验与 Pinterest 案例
  • English Title: MLE Core Cheatsheet: High-Frequency Q&A, Competitions & Pinterest
  • 分类:MLE · 标签:mle, interview-prep, cheatsheet, kaggle
  • 章节核心大纲:
    1. MLE 岗位核心八股与八股冲刺
    1. ML 竞赛 (Kaggle) 表现提升技巧
    1. Data Filtering (数据过滤) 与 Pinterest 实战案例

📄 mle-data-and-feature-engineering

  • 对应 aman_ai 源码文件:data-filtering.md, data-sampling.md, data-split.md, data-imbalance.md, preprocessing.md
  • 中文标题:MLE 数据与特征工程:质量管线、编码策略、不平衡处理、特征选择与漂移检测全景
  • English Title: MLE Data & Feature Engineering: Quality Pipelines, Categorical Encoding, Imbalance, Selection & Drift Detection
  • 分类:MLE · 标签:feature-engineering, data-quality, target-encoding, smote, feature-selection, drift-detection, imbalanced-learning, psi
  • 章节核心大纲:
    1. 数据质量与清洗
    1. 特征编码/变换/交叉
    1. 不平衡/特征选择/PSI 漂移监控

📄 mle-model-evaluation-and-debugging

  • 对应 aman_ai 源码文件:cross-validation.md, train-val-loss.md, debugging-dl-projects.md, evaluation-metrics.md
  • 中文标题:MLE 模型评估与调试工程:交叉验证策略、数据泄漏、偏差方差诊断、漂移检测与 A/B 验证
  • English Title: MLE Model Evaluation & Debugging Engineering: CV Strategies, Data Leakage, Bias-Variance Diagnosis, Drift & A/B Validation
  • 分类:MLE · 标签:evaluation, debugging, cross-validation, data-leakage, bias-variance, model-monitoring, drift-detection, ab-testing
  • 章节核心大纲:
    1. CV 策略与数据泄漏
    1. 高偏差 vs 高方差诊断
    1. 指标口径/漂移监控/A-B 验证

📄 mle-system-design-guide

  • 对应 aman_ai 源码文件:— (新增主题,待补充)
  • 中文标题:MLE 业务系统设计:推荐系统、搜索广告与风控架构全流程
  • English Title: MLE System Design Guide: Recommendation, Search & Risk Control
  • 分类:MLE · 标签:mle-system-design, recommendation, search-ads

📖 模块:Interviews/AIE

📄 aie-agent-systems-in-production

  • 对应 aman_ai 源码文件:agentic-design-patterns.md, loop-and-graph-engineering.md, context-engineering.md, agents.md
  • 中文标题:AIE Agent 生产系统:编排模式、上下文预算、可靠性工程与可观测性
  • English Title: AIE Agent Systems in Production: Orchestration Patterns, Context Budgeting, Reliability & Observability
  • 分类:AIE · 标签:agent, production, agent-orchestration, context-engineering, llm-ops, observability, tool-use, aie
  • 章节核心大纲:
    1. Agent 编排模式与循环工程
    1. 上下文工程与 token 预算
    1. 工具可靠性/成本/可观测性

📄 aie-core-cheatsheet

  • 对应 aman_ai 源码文件:prompt-engineering.md, personalize-LLMs.md, knowledge-graphs.md, knowledge.md
  • 中文标题:AIE 大模型系统工程师核心地图:SFT/LoRA/RAG/Agent 面试必考
  • English Title: AIE Core Cheatsheet: SFT, LoRA, RAG & Agent Interview Map
  • 分类:AIE · 标签:aie, llm-engineer, sft, lora, rag-interview
  • 章节核心大纲:
    1. Prompt Engineering 最佳实践 (CoT, Few-shot, System Prompt)
    1. Personalize LLMs (个性化 LLM 拟人化与用户画像)
    1. Knowledge Graphs 知识图谱与 Graph-RAG 结合

📄 aie-fine-tuning-and-alignment

  • 对应 aman_ai 源码文件:— (新增主题,待补充)
  • 中文标题:AIE 实战指南:企业级 SFT、LoRA 微调与 DPO/RLHF 偏好对齐落地
  • English Title: AIE Fine-Tuning Guide: Enterprise SFT, LoRA & DPO Alignment
  • 分类:AIE · 标签:sft, lora-merge, dpo-practical, grpo, alignment

📄 aie-system-design-guide

  • 对应 aman_ai 源码文件:— (新增主题,待补充)
  • 中文标题:AIE 大模型系统设计:千万级 RAG、Code Agent 与推理服务架构
  • English Title: AIE LLM System Design Guide: Production RAG, Agent & Serving
  • 分类:AIE · 标签:aie-system-design, production-rag, agent-infra, vllm-serving

📖 模块:Interviews/RS

📄 rs-core-cheatsheet

  • 对应 aman_ai 源码文件:top-30-papers.md, deep-rl.md
  • 中文标题:RS 核心知识地图:顶会必读 30 篇论文解构与深度强化学习 Deep RL
  • English Title: RS Core Cheatsheet: Top 30 Papers Breakdown & Deep RL
  • 分类:RS · 标签:research-scientist, top-papers, deep-rl, ppo, paper-breakdown
  • 章节核心大纲:
    1. Top 30 经典与 SOTA AI 论文解构
    1. Deep RL 核心算法 (Q-Learning, Policy Gradient, Actor-Critic)
    1. 白板推导规范与推导技巧

📄 rs-experiment-design-and-reproducibility

  • 对应 aman_ai 源码文件:hyperparameter-tuning.md, train-val-loss.md, commonErrors.md, tensorboard.md
  • 中文标题:RS 实验设计与可复现研究:消融实验、种子控制、超参搜索与可复现性全景全解
  • English Title: RS Experiment Design & Reproducible Research: Ablations, Seed Control, Hyperparameter Search & the Full Reproducibility Checklist
  • 分类:RS · 标签:experiment-design, reproducibility, ablation-study, hyperparameter-search, seed-control, tensorboard, paper-reproduction, research-methodology
  • 章节核心大纲:
    1. 实验设计原则与方差控制
    1. 消融设计与超参搜索
    1. 日志/复现清单与常见坑

📄 rs-math-proofs-and-derivations

  • 对应 aman_ai 源码文件:— (新增主题,待补充)
  • 中文标题:RS 数学推导与白板面经:PPO 剪切损失、DPO 闭式解与 RoPE 旋转矩阵
  • English Title: RS Math Proofs: PPO Clipped Loss, DPO Closed-Form & RoPE Matrix
  • 分类:RS · 标签:math-proofs, ppo-derivation, dpo-derivation, rope-proof

📄 rs-paper-deep-dive-framework

  • 对应 aman_ai 源码文件:— (新增主题,待补充)
  • 中文标题:RS 论文拆解与研究 Vision:如何向面试官复述 SOTA 论文创新点
  • English Title: RS Paper Deep Dive Framework: Articulating Novelty & Research Vision
  • 分类:RS · 标签:paper-deep-dive, research-vision, deepseek-r1, academic-taste

📖 模块:Interviews/DS

📄 ds-ab-testing-case-studies

  • 对应 aman_ai 源码文件:— (新增主题,待补充)
  • 中文标题:DS A/B 测试 Case Studies:CUPED 方差降低、SRM 检查与商业归因
  • English Title: DS A/B Testing Case Studies: CUPED, SRM Checks & Attribution
  • 分类:DS · 标签:ab-testing-cases, cuped, srm, attribution

📄 ds-causal-inference-and-experimentation

  • 对应 aman_ai 源码文件:— (新增主题,待补充)
  • 中文标题:DS 因果推断实战:PSM 倾向得分匹配、DiD 双重差分与 Synthetic Control
  • English Title: DS Causal Inference: PSM, Difference-in-Differences & Synthetic Control
  • 分类:DS · 标签:causal-inference, psm, did, synthetic-control, iv-2sls

📄 ds-core-cheatsheet

  • 对应 aman_ai 源码文件:online-testing.md, experimentation-biases.md, drift.md, causalInference.md
  • 中文标题:DS 核心知识地图:因果推断、A/B 测试与 Data Drift
  • English Title: DS Core Cheatsheet: Causal Inference, A/B Testing & Drift
  • 分类:DS · 标签:ds, data-scientist, causal-inference, ab-testing, psi-drift
  • 章节核心大纲:
    1. 因果推断 (Causal Inference) PSM/DiD/IV 理论
    1. 在线 A/B 测试、CUPED 方差降低与 Sample Ratio Mismatch (SRM)
    1. 实验偏差防护 (Selection Bias, Novelty Effect) 与 Data Drift 监测

📄 ds-statistics-and-experiment-design

  • 对应 aman_ai 源码文件:online-testing.md, experimentation-biases.md, drift.md, probability-calibration.md
  • 中文标题:DS 统计与实验设计:假设检验、功效分析、样本量计算、多重比较、SRM 与 CUPED 全景全解
  • English Title: DS Statistics & Experiment Design: Hypothesis Testing, Sample Size, Multiple Comparisons, SRM & CUPED
  • 分类:DS · 标签:statistics, experiment-design, ab-testing, hypothesis-testing, sample-size, multiple-comparisons, srm, cuped
  • 章节核心大纲:
    1. 功效分析与样本量计算
    1. 多重比较/p-hacking/peeking
    1. SRM/AA 测试/CUPED 方差缩减

🧠 深入探索 TalentMe 全景技术图谱与备考路线

本文选自 TalentMe AI 技术专栏与高维职业罗盘。支持双模态 Obsidian 本地私域同步、艾宾浩斯智能复习与 IDE 内嵌 AI 导师模拟面试。

👉 访问 TalentMe 技术专栏 →


Discover more from AirSOTA – Air School Of Thoughts AtoZ

Subscribe to get the latest posts sent to your email.