From Machine Learning to Generative AI: A Complete Open-Source AI Learning Course

This article introduces an open-source AI course developed by FIAP MBA professor Ahirton Lopes, covering a complete knowledge system from supervised/unsupervised machine learning to deep learning and generative AI. The course includes 15 practical demonstration notebooks using mainstream tools like Python, TensorFlow, and Keras, suitable for learners who want to systematically study artificial intelligence.

• Original Author/Maintainer: Prof. Dr. Ahirton Lopes (FIAP MBA Professor)
• Source Platform: GitHub
• Original Title: AI-Foundation-and-Learning-Models
• Original Link: https://github.com/ahirtonlopes/AI-Foundation-and-Learning-Models
• Release Date: May 2026

---

In today's era of rapid development of artificial intelligence technology, systematic learning of AI has become an urgent need for many technical practitioners and students. However, learning resources on the market are often scattered, fragmented, or too theoretical and lack practicality. This open-source course developed by FIAP MBA professor Ahirton Lopes just fills this gap.

The biggest features of this course are its completeness and practicality. It is not just a theoretical lecture note, but a complete learning path containing 15 runnable Jupyter Notebooks, covering the full spectrum of content from traditional machine learning to cutting-edge generative AI technologies. Learners can learn while practicing, deepening their understanding through actual code.

---

The entire course system is divided into three major modules, forming a step-by-step learning ladder:

This module lays a solid AI foundation for beginners, including 5 core demos:

Demo1: URL Classifier (Introduction to Supervised Learning)

As the first demo of the course, this demo introduces the basic concepts of supervised learning. Learners will build a URL classifier, learn how to extract features from text data, and how to train and evaluate a basic classification model. This lays the conceptual foundation for more complex models later.

Demo2: Comparison of Model Evaluation Metrics

Is Accuracy sufficient? This demo deeply explores various evaluation metrics in classification tasks, including F1 score, AUC-ROC curve, and confusion matrix. Learners will understand why accuracy may be misleading in some scenarios and how to choose appropriate evaluation metrics.

Demo3: Call Center Demand Forecasting (Time Series)

Time series forecasting is an important branch of machine learning. This demo uses real call center data to show how to model and forecast time series data, covering key techniques such as seasonal analysis and trend extraction.

Demo4: Fraud Detection (Imbalanced Data Handling)

Financial fraud detection is a classic application of machine learning in industry. This demo focuses on handling extremely imbalanced datasets, introducing techniques such as oversampling, undersampling, and cost-sensitive learning to help learners understand how to apply machine learning in real business scenarios.

Demo5: K-Means Clustering on Crime Data (Unsupervised Learning)

Unsupervised learning is a powerful tool for discovering hidden patterns in data. This demo uses the K-Means algorithm to perform clustering analysis on crime data, showing how to discover the internal structure of data without labels.

After laying the foundation, the course enters the field of deep learning, including 5 core demos:

Demo6: Implementing Gradient Descent and Backpropagation from Scratch

This is a key lesson for understanding the principles of deep learning. Learners will implement gradient descent algorithm and backpropagation mechanism purely using NumPy without any deep learning framework. This "first principles" learning approach helps learners truly understand how neural networks work.

Demo7: Autoencoders and Latent Space

Autoencoders are a special type of neural network structure used to learn effective representations (encoding) of data. This demo shows how to build an autoencoder, understand the concept of latent space, and how to apply autoencoders to dimensionality reduction and feature learning.

Demo8: CIFAR-10 Image Classification (Convolutional Neural Networks)

Convolutional Neural Networks (CNNs) are the cornerstone of computer vision. This demo uses the classic CIFAR-10 dataset to teach step-by-step how to build and train a CNN model, covering core components such as convolutional layers, pooling layers, and batch normalization.

Demo9: Spam SMS Filtering (Recurrent Neural Networks)

Recurrent Neural Networks (RNNs) are the basic architecture for processing sequence data. This demo builds a spam SMS classifier, showing how RNNs capture temporal dependencies in text.

Demo10: IMDB Sentiment Analysis (LSTM)

Long Short-Term Memory (LSTM) networks solve the gradient vanishing problem of traditional RNNs. This demo uses the IMDB movie review dataset to show the powerful capabilities of LSTMs in sentiment analysis tasks, while introducing the concept of word embedding.

The final module of the course focuses on the currently hottest generative AI technologies:

Demo11: CIFAR-10 Image Generation (GAN)

Generative Adversarial Networks (GANs) are the pioneering work of generative models. This demo shows how to build a basic GAN, understand the game process between the generator and discriminator, and the unique challenges of training GANs.

Demo12: Attention Mechanism and Transformer Architecture

The Transformer architecture is the foundation of modern large language models. This demo deeply explains the principles of the attention mechanism, shows how self-attention works, and how Transformers have revolutionized the field of natural language processing.

Demo13: Stable Diffusion Text-to-Image Generation

Stable Diffusion is currently the most popular open-source text-to-image generation model. This demo shows how to use and optimize Stable Diffusion, understand the basic principles of diffusion models, and how to perform prompt engineering.

Demo14: LangChain Large Language Model Orchestration

LangChain is a mainstream framework for building LLM applications. This demo introduces core concepts such as chains, agents, and memory, showing how to integrate large language models into practical applications.

Demo15: YOLO Real-Time Object Detection

YOLO (You Only Look Once) is a classic algorithm for real-time object detection. This demo shows how to run the YOLO model for real-time object recognition and understand the design philosophy of single-stage detectors.

---

The course uses mainstream industry technology stacks to ensure that the skills learners master have practical application value: