MuXinCG's Blog

Back

Roadmap

Computer Science Learning Map

Course notes, algorithm reviews, and problem writeups arranged into practical reading tracks.

25

Published posts

5

Series

29

Topic tags

CS229

Machine Learning Foundations

Regression, classification, generalized linear models, and the mathematical base for later model work.

4 notes Updated March 2, 2026 machine learning cs229 notes
  1. 01 CS229 Machine Learning CS229 Machine Learning - Introduction and Overview CS229 的监督、无监督与强化学习三大板块,所需的线性代数、概率论与微积分基础,以及本系列的整体规划。 machine learning cs229 notes
  2. 02 CS229 Machine Learning CS229 Machine Learning - Linear Regression 把误差项假设为高斯噪声后,最大似然估计恰好等价于最小二乘,再用梯度下降迭代求解,并以局部加权回归拟合非线性。 machine learning cs229 notes
  3. 03 CS229 Machine Learning CS229 Machine Learning - Logistic Regression 分类问题的目标变量只取离散值,用 sigmoid 把线性输出压到 (0,1) 区间,即得到逻辑回归模型。 machine learning cs229 notes
  4. 04 CS229 Machine Learning CS229 Machine Learning - Generalized Linear Models 指数族的统一形式 p(y;η)=b(y)exp(ηᵀT(y)−a(η)),由它构造广义线性模型可把最小二乘与逻辑回归纳入同一框架。 machine learning cs229 notes

CS336

Language Modeling From Scratch

Tokenizer, Transformer, systems, scaling laws, data processing, alignment, and reasoning RL.

  1. 01 CS336 Language Modeling from Scratch CS336 Language Modeling from Scratch - Introduction CS336 几乎不提供脚手架代码,要求从 tokenizer、Transformer 一路手写到分布式训练与对齐,本系列按 5 个 Assignment 组织。 language models cs336 notes
  2. 02 CS336 Language Modeling from Scratch CS336 - Assignment 1: Tokenizer & Transformer Basics 只用 PyTorch primitives 手写 BPE 分词器、RMSNorm/RoPE/SwiGLU 与 AdamW,拼出可在 TinyStories 上训练收敛的 decoder-only 模型。 language models cs336 transformer tokenization
  3. 03 CS336 Language Modeling from Scratch CS336 - Assignment 2: Systems (Triton & Distributed) 先 profile 定位瓶颈,用 Triton 手写 FlashAttention-2 以分块加 online softmax 避开 N×N 注意力矩阵,再实现 DDP 与优化器状态分片。 language models cs336 triton distributed training
  4. 04 CS336 Language Modeling from Scratch CS336 - Assignment 3: Scaling Laws 训练一批小模型拟合幂律再外推,用 IsoFLOP 方法在固定算力预算下寻找 loss 最低点,预测 compute-optimal 的模型规模与数据量。 language models cs336 scaling laws notes
  5. 05 CS336 Language Modeling from Scratch CS336 - Assignment 4: Data Processing & Filtering 把原始 Common Crawl 变成预训练语料:WARC 抽正文、fastText 识别语言、Gopher 规则与分类器过滤、PII 脱敏与去重。 language models cs336 data notes
  6. 06 CS336 Language Modeling from Scratch CS336 - Assignment 5: Alignment & Reasoning RL SFT 只在 response 上计算 loss 做指令微调,再用 expert iteration 与 GRPO 强化学习提升 MATH 等数学推理基准表现。 language models cs336 alignment reinforcement learning

Algorithms

Algorithm System

Core techniques, data structures, graph theory, strings, math, geometry, and review notes.

  1. 01 Algorithm Miscellany Algorithm Miscellany - Basics 用 bitset 压位加速集合运算,树上前缀和与差分借 LCA 处理路径统计,倍增法压缩大状态空间的递推。 algorithm review
  2. 02 Algorithm Miscellany Algorithm Miscellany - Prefix Sum and Difference 前缀和把区间求和降到 O(1),差分把区间修改降到 O(1),二维情形用容斥原理查询任意子矩阵。 algorithm data structure review
  3. 03 Algorithm Miscellany Algorithm Miscellany - Search A* 靠不超过真实距离的启发函数保证最优性,迭代加深与 IDA* 分别用深度和代价上限控制 DFS,配合记忆化与可行性剪枝。 algorithm search review
  4. 04 Algorithm Miscellany Algorithm Miscellany - Dynamic Programming 逐维求和把 k 维前缀和做到 O(kN),据此得到 O(n·2ⁿ) 的子集和 DP,并用状态合法性判定重构最长不下降子序列。 algorithm dynamic programming review
  5. 05 Algorithm Miscellany Algorithm Miscellany - String 字符串哈希由哈希值不等推出原串不等,Trie 按字符转移建树,KMP 用前缀函数记录最长相等真前后缀以避免回溯。 algorithm string review
  6. 06 Algorithm Miscellany Algorithm Miscellany - Mathematics 数论中整除与约数的基本性质,以及多项式与生成函数的基本概念。 algorithm math review
  7. 07 Algorithm Miscellany Algorithm Miscellany - Data Structures 并查集用路径压缩加按大小合并维护集合归属,再以分块、ST 表与线段树支撑区间查询与修改。 algorithm data structure review
  8. 08 Algorithm Miscellany Algorithm Miscellany - Graph Theory 树的中心使最长链最短、重心使最大连通块不超过一半,重链剖分把任意路径拆成 O(log n) 条连续链交给线段树维护。 algorithm graph theory tree review
  9. 09 Algorithm Miscellany Algorithm Miscellany - Computational Geometry Algorithm Miscellany series: computational geometry-related algorithms. algorithm computational geometry review
  10. 10 Algorithm Miscellany Algorithm Miscellany - Miscellaneous 离散化用排序去重加二分把值域压到 O(n),CDQ 分治则把动态问题转化为静态的点对问题。 algorithm review

OJ Practice

Problem Solving

Selected Codeforces and Luogu writeups focused on construction, bit tricks, trees, and implementation details.

  1. 01 Codeforces Solutions XOR Convenience - Codeforces Solution 把所有异或结果都导向最后一个位置,构造出每个值与其后某个值异或后等于其下标的排列。 algorithm codeforces construction bit manipulation
  2. 02 Luogu Solutions P1131 [ZJOI2007] Time Synchronization - Luogu Solution 树形 DP 求每棵子树内根到叶的最长路径,再在各子节点分支上补齐差值,只增加边权即可让所有叶子到达时间一致。 algorithm luogu tree dp tree
  3. 03 Codeforces Solutions MEX Reordering - Codeforces Solution 通过分析 0 与 1 的出现次数构造重排,使任意前缀与后缀的 MEX 都不相等,单次遍历 O(n) 完成。 algorithm codeforces construction
  4. 04 Codeforces Solutions The Curse of the Frog - Codeforces Solution 青蛙在数轴上跳跃且每种跳法有周期性罚时,贪心优先使用无罚时的跳法,O(n) 求出到达目标的最少罚时次数。 algorithm codeforces greedy
  5. 05 Codeforces Solutions XOR Array - Codeforces Solution 用前缀异或配合唯一重复值构造正整数数组,使恰好指定的子数组异或和为零,边算边输出,空间 O(1)。 algorithm codeforces construction bit manipulation