Write a Simple Neural Network and Train It — No PhD Required

 

Let’s not try to build GPT-4 today. Let’s build something simple. Something bite-sized. A little neural network that learns, fails, improves, and eventually gets kind of smart. No magic. No math degree. Just a laptop, Python, and a good attitude.

Wait… What Is a Neural Network?

It’s just a pile of weighted numbers that guess stuff. You give it some input (like pixel values or numbers), it runs them through math layers (like passing hot gossip through coworkers), and it spits out a prediction. If it’s wrong, it updates its weights, like saying, “Ah, maybe Susan wasn’t the one who said that thing.” Eventually, with enough tries, it gets better. That’s learning.

Teach a Network to Recognize Handwritten Digits

We’re going to train a neural network on the classic MNIST dataset. It’s just a bunch of 28x28 pixel images of handwritten digits (0–9). Like flashcards for AI toddlers.

Step 1: Set Up Your Tools

You’ll need:

pip install torch torchvision matplotlib

That’s it. Three things:

• torch: PyTorch, the deep learning library we’ll use
• torchvision: To get the MNIST dataset easily
• matplotlib: For making things pretty

Step 2: The Code — A Barebones Neural Network

What Just Happened?

Let’s decode this like a human:

• We got data: MNIST is the go-to dataset for “AI babies” because it’s small, simple, and easy to learn from. 60,000 grayscale images of numbers.
• We built a tiny brain: one hidden layer with 128 neurons. It flattens the image (28x28 = 784 pixels), feeds them into a layer, ReLU’s them (a kind of “on/off” filter), and outputs 10 guesses (0–9).
• We trained it: like doing flashcards. Show image → guess → check answer → update model.

After five epochs, your little AI toddler is already reading digits better than a distracted first grader.

But Where’s the Magic?

There’s no magic. Just a feedback loop:

1. Guess.
2. Measure how bad the guess is (loss).
3. Adjust weights to make future guesses better (backpropagation).
4. Repeat until the model stops sucking.

That’s the entire game. Fancy architectures just do this a billion times with way more layers and data.

Tips From the Trenches

• Start small. One hidden layer is fine. No shame in basic brains.
• Visualize the progress. Plot the loss over time or test accuracy. It’s motivational.
• Break stuff on purpose. Change the learning rate. Remove ReLU. Watch it fail. That’s how you learn.

Why This Matters

You just made a neural network. Not just used one — built one. You saw how it works: not magic, just math and persistence.

Yes, you’ll hear about convolutional layers, transformers, LLMs, and attention heads. But underneath? It’s all the same idea.

Inputs — layers — guess — loss — fix — repeat.

This is the plumbing behind the future. You don’t need to be a wizard — just someone curious enough to build a thing, run it, break it, and try again.