Artificial Intelligence and Machine Learning Professor
I taught about the fundamentals of artificial intelligence and machine learning, neural networks, convolutional neural networks, recurrent neural networks, generative adversarial networks, and reinforcement learning.
Teaching AI by building it
In 2024, everyone was talking about ChatGPT and AI. Most people did not understand how it worked.
From February to September 2024, I taught Artificial Intelligence and Machine Learning at Andrés Bello Catholic University in Caracas. It was a part time role, and it lasted eight months. My goal was simple. I wanted students to learn how AI really works, not only how to use it. So instead of only calling APIs, we built things from scratch. That way they could see what happens inside.
One project every week
I built the course around projects. Every week, the class ended with something that worked. In total, we made more than 10 projects. Students used their programming skills on real problems: movie recommendations, image classification, background removal, audio denoising, voice to voice chatbots, automated trading and video games.
I also changed how the theory classes worked. I gave students notes with blanks to complete during the lesson, and I ran quick opinion polls. Both forced them to pay attention. They also made students feel rewarded when they discovered they had really learned something new.
Starting with basic machine learning
We started with a simple question. How does a machine know if a background is light or dark?
Instead of only explaining it, we built ChromAI. We used Logistic Regression to teach a computer to detect contrast. Logistic Regression is one of the simplest models. It learns to choose between two answers, here light or dark. The project was simple, but it helped students understand the basics.
Then a student asked how Netflix recommends shows. That question led to ScreenBuddy. We used KNeighbors, a method that finds the most similar examples, to build a recommendation system. Students learned how computers find patterns in what people like.
Building neural networks by hand
Once students felt comfortable, they wanted more. One student said he wanted to build a neural network from scratch, without a library doing the work.
So we did. We rebuilt AlexNet, a famous computer vision model. We wrote each layer of the network by hand. It failed a lot at first. After we adjusted the settings, it learned to identify flowers.
Training took too long, so we learned about Transfer Learning with MobileNetV2. Transfer Learning means you start from a model that already learned from other images, and you only teach it the last part. It saved us a lot of time.
After that, we tried more advanced projects. We used the Segment Anything Model (SAM) to build RemoveBG, which removes the background from images automatically. We also used Autoencoders to make low quality images look better by filling in missing details. An autoencoder is a network that learns to compress an image and then rebuild it.
Sound and data over time
Next, we worked with audio. We built Denoiser with SepFormer. It removes background noise from audio and makes speech clearer.
Then students wanted to predict Bitcoin prices. We used Recurrent Neural Networks (RNNs, LSTMs and GRUs) with data from Yahoo Finance. These networks read data in order, so they are good for values that change over time. One student’s model predicted Bitcoin would reach $72,000. It really happened.
Voices, translation and rockets
Near the end of the semester, students wanted to try harder things. One asked about building an AI that could hold a conversation.
We built AIFriend by combining Text to Speech (Coqui-TTS) with Speech to Text. It could hold voice conversations in real time. Students saw how modern AI assistants work.
We also built a Transformer from scratch to translate Spanish to English. A Transformer is the type of model behind modern language models. Students coded its Self-Attention mechanism themselves. Self-Attention lets each word look at the other words in the sentence to understand its meaning.
We also learned to deploy third party models with tools like Cloudflare AI Workers and Azure ML Studio. That class showed students how to put a model online so other people can use it.
For the final project, we used Reinforcement Learning to train an AI to land a lunar rover. In Reinforcement Learning, the AI learns by trial and error, with rewards for good moves. It crashed hundreds of times. Then it learned to land safely.
From fear to confidence
The best part was not the projects. It was watching students gain confidence.
At the start, many were scared of terms like “Neural Network.” By the end, they were discussing which activation functions work best, and why.
Students rated the course 5.91 out of 6.0. What mattered more is that students got internships and jobs, because they could show recruiters real projects they built themselves. They could explain how attention mechanisms work, because they had coded them.
That was the goal from the first class. Students did not only use AI. They understood it, because they built it.
Here are the projects we built during the course:
ChromAI Class 1. ChromeAI is an AI that determines the optimal font color (light or dark) based on the background color, improving readability in web interfaces. It uses machine learning techniques to predict whether text should be white or black based on the provided background color.
ScreenBuddy Class 2. ScreenBuddy is an artificial intelligence app that recommends movies based on the preferences of other users with similar likes. The project combines theory and practice in recommendation systems, machine learning, and data analysis.
Identity Class 3. Identity is an artificial intelligence project that implements a facial unlocking system. Given a person's image, the model is able to identify who it is and, if appropriate, unlock the device. This approach seeks to simplify device access, avoiding traditional methods such as PINs or patterns.
Alexnet Class 4. This project focuses on developing an artificial intelligence capable of identifying elements in images. Through a series of practical implementations, different convolutional neural network (CNN) architectures are explored.
RemoveBG Class 5. This project aims to develop an Artificial Intelligence model capable of automatically removing backgrounds from images.
Upscaling Class 6. This repository contains a Jupyter Notebook that demonstrates how to use pre-trained AI models Hugging Face and Replicate to perform image super-resolution. The goal is to upsample a small image (e.g., 96x96) to a larger one (e.g., 512x512) by improving its quality, rather than simply scaling the pixels.
Denoiser Class 7. This project is the result of a class on audio processing. The main objective is to develop an artificial intelligence capable of removing background noise from an audio recording.
Bitcoin price predictions Class 8. This project uses a recurrent neural network (RNN) to analyze the price of Bitcoin and predict its fluctuations.
Spanish to English Translator Class 9. This project implements an artificial intelligence model to translate text from Spanish to English, using Natural Language Processing (NLP) techniques and a transformer. It also includes a web interface for interacting with the translator.
Usage of third party models Class 10. This repository contains a Jupyter Notebook tutorial that demonstrates how to deploy and use third-party AI models using tools such as Cloudflare AI Workers or Azure ML Studio.
AIFriend Class 11. AIFriend is an artificial intelligence voice-to-voice designed to converse with you. You can speak to it and it will respond in a human voice, simulating a conversation with a friend. This project was born as a solution to the growing lack of social interaction in an increasingly digital world.
Lunar Rover Landing Class 12. This project focuses on training a reinforcement learning agent to land a rover on the moon safely and efficiently.