Neural Networks

Explore & Discover

Neural networks are the engine behind face ID on your phone, the recommendations on your Netflix, and every chatbot you've ever talked to. Start by watching 3Blue1Brown's free YouTube series "But what is a neural network?" — it uses beautiful animations to show how they actually work without needing any math background. Then explore Google's free Teachable Machine (teachablemachine.withgoogle.com) and train a model using your webcam to recognize hand gestures — no coding required. Notice how more training examples make the model smarter. Researchers at the University of Utah work on neural networks right here on the Wasatch Front. You're ready for the next step when you can explain what a neuron does in a neural network and why training data matters.

Learn the Basics

Learn the core concepts that make neural networks tick using free resources. Fast.ai's "Practical Deep Learning for Coders" (free at fast.ai) starts with working code and explains the theory as you go — no PhD required. Understand these key ideas: layers (input, hidden, output), weights and biases, activation functions, and backpropagation. Khan Academy's "Neural Networks" section explains backpropagation with clear visuals. The key insight: a neural network learns by making predictions, measuring how wrong it was (the loss), and adjusting its weights slightly to be less wrong next time — thousands of times over. You're ready for the next step when you can draw a simple neural network diagram and label its layers, weights, and activation functions.

Build Your First Project

Build your first real neural network using Python and TensorFlow or PyTorch — both are free. If you haven't coded before, start with Google Colab (free, runs in your browser, no installation needed). Follow the official TensorFlow beginner tutorial that trains a network to recognize handwritten digits from the MNIST dataset — it's the "Hello World" of deep learning. You'll write about 20 lines of code to build, train, and test a model that gets 97%+ accuracy. Walk through each line and understand what it does before running it. You're ready for the next step when you have a working MNIST digit classifier running in Google Colab that achieves at least 95% accuracy on the test set.

Experiment & Iterate

Experiment with your neural network to understand how each piece affects performance. Try changing the number of hidden layers or neurons and observe what happens to accuracy. Deliberately overfit your model (make it too complex for the data) and then fix it by adding dropout layers — a regularization trick that randomly turns off neurons during training to prevent memorization. Change the learning rate and see how it affects training speed and final accuracy. Use Weights & Biases (free account) or simple matplotlib graphs to visualize your training curves. Keeping a lab notebook of your experiments is exactly what AI researchers do. You're ready for the next step when you can show experiment results comparing at least three different network architectures and explain which performed best and why.

Advanced Techniques

Move into convolutional neural networks (CNNs) — the architecture that made image recognition take off. A CNN uses special layers that scan for patterns like edges, shapes, and textures, which is way more efficient than a plain neural network for images. Follow the free fast.ai lesson on CNNs or use TensorFlow's official image classification tutorial. Train a CNN to classify images in a dataset you care about — Kaggle has hundreds of free datasets on everything from dog breeds to satellite imagery of Utah landscapes. Learn transfer learning: taking a pre-trained model like ResNet and fine-tuning it on your own data with just a few lines of code. You're ready for the next step when you can train a CNN using transfer learning that classifies images into at least three categories with over 85% accuracy.

Final Project Showcase

Design and train a neural network that solves a problem you actually care about. It could be a Utah wildlife image classifier, a model that predicts your gaming performance, or a digit recognizer you deploy as a simple web app using Gradio (free, one page of code). Document your full process: problem statement, dataset source, architecture choices, training results, and what you'd improve with more time. Share your project on GitHub (free) and post it in an AI community like fast.ai's forums or r/MachineLearning. Write a short explanation of your model that a non-technical friend could understand. You're ready for the next step when you can demo your working model, explain why you chose your architecture, and describe what the training curve told you about your model's learning.