First: Understand What You're Actually Trying to Learn
These three terms are used interchangeably and incorrectly all the time:
| Term | What It Means | Example |
|---|---|---|
| AI | Any technique that makes machines appear intelligent | Chess engines, recommendation systems, voice assistants |
| Machine Learning | AI that learns patterns from data without being explicitly programmed | Spam filters, image classifiers, price prediction |
| Deep Learning | ML using multi-layer neural networks — mimics the brain loosely | ChatGPT, image generation, speech recognition |
| Data Science | Using statistics and code to find patterns and insights in data | Business analytics, A/B testing, dashboards |
When most people say "I want to learn AI," they mean machine learning and deep learning. That is what this roadmap covers.
The 6-Phase AI Learning Roadmap
Phase 1 — Python (Weeks 1–6)
You cannot learn AI without Python. It is the universal language of AI and machine learning. Before touching any AI library, you must be comfortable with: variables, data types, functions, loops, list comprehensions, dictionaries, classes, and importing libraries.
Do not skip this. Students who try to jump to TensorFlow or PyTorch without Python fluency hit a wall within days and give up. Four to six weeks of daily Python practice makes everything else 5× faster.
Phase 2 — Maths Foundations (Weeks 4–8, parallel with Python)
You need enough maths to understand what AI algorithms are doing — not how to derive them. Focus on three areas:
- Statistics: Mean, median, standard deviation, probability, distributions, correlation
- Linear Algebra: Vectors, matrices, matrix multiplication, dot products
- Calculus (conceptual only): What a derivative is, what minimising a function means
You do NOT need to derive backpropagation by hand. You need to understand why gradient descent works conceptually. Khan Academy covers all three areas for free.
Phase 3 — Machine Learning with scikit-learn (Weeks 7–14)
scikit-learn is Python's beginner-friendly ML library. Start here before PyTorch or TensorFlow. Learn in this order:
- Supervised learning: linear regression, logistic regression
- Decision trees and random forests
- Train/test splits, cross-validation
- Overfitting, underfitting, bias-variance tradeoff
- Model evaluation: accuracy, precision, recall, F1 score
from sklearn.datasets import load_iris
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split
X, y = load_iris(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)
model = RandomForestClassifier()
model.fit(X_train, y_train)
print("Accuracy:", model.score(X_test, y_test))That is a complete ML model in 8 lines. Start there.
Phase 4 — Deep Learning with PyTorch (Weeks 14–22)
PyTorch is now the dominant deep learning framework in both research and production. Learn:
- Tensors and how they differ from NumPy arrays
- Building a neural network with torch.nn
- Forward pass, loss function, backpropagation
- Convolutional neural networks (CNNs) for images
- Transfer learning — using pretrained models
Phase 5 — Build 3–5 Real AI Projects (Weeks 18–26)
This is the most important phase. Projects turn skills into a portfolio. Build at minimum:
- An image classifier (train on any dataset you care about)
- A sentiment analysis model (classify positive/negative text)
- A price or outcome prediction model with real data
- One project that uses a pre-trained LLM via API (e.g., OpenAI, Gemini, or Mistral)
Upload all of them to GitHub. Write a clear README for each. This is what employers and university evaluators actually look at.
Phase 6 — Specialise and Go Deeper (Month 6+)
After the first five phases, you have a foundation. Now choose a direction:
- NLP / LLMs: HuggingFace Transformers, fine-tuning, RAG systems
- Computer Vision: Object detection, segmentation, OpenCV, YOLO
- MLOps: Model deployment, Docker, cloud platforms (AWS, GCP)
- Reinforcement Learning: Game-playing agents, reward modelling
- Data Science path: Statistical analysis, A/B testing, business intelligence
Realistic Time Expectations
| Milestone | Time (1–2 hrs/day) | What You Can Do |
|---|---|---|
| Python basics complete | 4–6 weeks | Write scripts, automate tasks, use libraries |
| First ML model | 10–12 weeks | Train classifiers, understand predictions |
| First deep learning model | 16–20 weeks | Image classification, basic neural networks |
| Portfolio with 3+ projects | 20–26 weeks | Apply for internships or entry-level roles |
| Job-ready (entry level) | 9–12 months | Full ML pipeline, deployment, real-world experience |
Best Free Resources to Learn AI in 2026
| Resource | What It Covers | Cost |
|---|---|---|
| BitWithBite AI & ML Course | Python to ML to deep learning, project-first | Free trial |
| Fast.ai Practical Deep Learning | Deep learning, PyTorch, computer vision, NLP | Free |
| Google ML Crash Course | ML fundamentals, TensorFlow basics | Free |
| Andrew Ng ML Specialization (Coursera) | Classical ML, deep learning, MLOps | Auditable free |
| Kaggle Learn | Python, ML, deep learning, NLP, CV | Free |
| HuggingFace Course | Transformers, NLP, LLMs, fine-tuning | Free |
| Khan Academy | Statistics, linear algebra, calculus | Free |
Do You Actually Need Maths for AI?
This is the question that puts the most people off starting. The honest answer: you need conceptual maths, not computational maths.
You need to understand that gradient descent minimises a loss function by moving in the direction of steepest descent. You do not need to derive that gradient by hand. PyTorch does that for you automatically with autograd.
You need to understand that a matrix multiplication combines feature representations. You do not need to compute 1000×1000 matrix products by hand. NumPy does that in milliseconds.
🔑 The Real Prerequisite for AI
Curiosity and persistence matter far more than prior maths ability. The mathematical intuition you need develops naturally as you build models and observe their behaviour — it doesn't have to come first. Start building; the maths will follow.
Frequently Asked Questions
📚 Related Articles
Irfana built BitWithBite to make AI and programming education genuinely accessible — cutting the fluff and focusing on practical, project-based learning. She has guided hundreds of students through their first AI models.