Learn AI from Scratch: 4 Stages to Save Two Years of Mistakes

01 | Clarify What AI Actually Covers

Many newcomers confuse whether to start with machine learning or deep learning without understanding their relationship. A hierarchical diagram is used to show four layers: Foundation (linear algebra, probability, calculus, Python), Machine‑Learning (supervised/unsupervised learning, classic algorithms such as decision trees and SVM), Deep‑Learning (neural networks, CNN, RNN, Transformer), and Application (large models like GPT/LLM, computer vision, NLP). The analogy is that math and Python are the foundation, machine learning is the framework, deep learning is the superstructure, and applications are the finishing touches.

02 | Four Stages from Zero to Practice

Stage 1 — Lay the Foundation (1‑2 months)

Goal: be able to write Python code and read basic math formulas.

Math : focus on three essential topics—linear algebra (matrix operations, eigen‑decomposition), probability & statistics (Bayes theorem, hypothesis testing), and calculus (partial derivatives, chain rule, gradient descent).

Programming : learn Python syntax, NumPy, Pandas, and Matplotlib. Python is the dominant language for AI; other languages are not recommended.

Recommended learning method: “learn while doing.” Use a simple Kaggle dataset (e.g., Titanic survival) to practice data loading, analysis, and visualization.

Stage 2 — Machine‑Learning Basics (2‑3 months)

This stage builds the AI mindset by covering core concepts such as train/validation/test splits, over‑/under‑fitting, loss functions, optimizers, and evaluation metrics (accuracy, precision, recall, F1). Classic algorithms are introduced in order: linear regression, logistic regression, decision trees & random forests, SVM, and K‑means clustering. The Scikit‑learn library is sufficient to implement these models with a few lines of code. Practical advice: complete Kaggle beginner competitions (Titanic classification, house‑price regression) to run the full pipeline—data cleaning → feature engineering → model training → hyperparameter tuning → submission.

Stage 3 — Deep‑Learning Advancement (2‑3 months)

Key steps:

Understand neural networks: forward and backward propagation, activation functions (ReLU, Sigmoid, Tanh), loss functions (cross‑entropy, MSE), and optimizers (SGD, Adam).

Study CNNs for image tasks—convolution, pooling, receptive field, and classic architectures like ResNet and VGG.

Learn RNN/LSTM for sequential data and the Transformer architecture, emphasizing the attention mechanism.

Start using PyTorch (preferred over TensorFlow for its intuitiveness) to build, train, save, and load models, and run experiments on free GPUs such as Google Colab.

Suggested projects: image classification on CIFAR‑10 or a cat‑dog dataset, sentiment analysis with LSTM, and a small self‑implemented Transformer.

Stage 4 — Large‑Model Era Practice (1‑2 months)

Focus on the fastest‑growing area in 2024. Core topics include:

LLM fundamentals (pre‑training → instruction fine‑tuning → RLHF).

Prompt engineering.

Retrieval‑Augmented Generation (RAG).

Parameter‑efficient fine‑tuning methods such as LoRA and QLoRA.

Agent development with LangChain or LlamaIndex.

Tools: Hugging Face Transformers, LangChain/LlamaIndex, and Ollama for local model serving. Recommended hands‑on projects: build a RAG‑based Q&A system, fine‑tune a small model with LoRA, and create a simple AI agent that searches, summarizes, and generates reports.

03 | Curated Learning Resources

Only a few high‑quality resources are listed:

Math : 3Blue1Brown linear‑algebra series, MIT 18.06 OCW.

Python : Liao Xuefeng’s tutorial, "Python Data Science Handbook" (NumPy, Pandas, Matplotlib).

Machine Learning : Andrew Ng’s Coursera course, Zhou Zhihua’s "Machine Learning" textbook, Li Hongyi’s B‑site lectures.

Deep Learning : Andrew Ng’s Deep Learning Specialization, Li Hongyi’s deep‑learning videos, "Dive into Deep Learning" by Li Mu.

Large Models : Hugging Face NLP course, LangChain documentation, Andrej Karpathy’s YouTube channel.

04 | Five Common Pitfalls to Avoid

Buying many courses but not completing any—focus on one course and finish it.

Only watching videos without coding—implement at least a small snippet for every new concept.

Striving for perfect math before coding—understand enough to follow formulas, fill gaps later.

Chasing every new paper—master classic models first, then explore new research.

Studying alone—join communities (Kaggle, Zhihu AI, GitHub Discussions) for peer support.

Conclusion

The roadmap condenses six to ten months of focused study into four stages: foundation (math + Python), machine learning (classic algorithms + Scikit‑learn), deep learning (neural networks + PyTorch), and large‑model practice (LLM + RAG + Agents). Following this path can reduce the learning time from years of trial‑and‑error to under a year and produce portfolio‑ready AI projects.