I Built a Neural Network from Scratch on My Phone — No Laptop, No Frameworks
I'm a high school student in Ghana. I don't have a laptop. Everything I code, I do on my phone using Google Colab.
Yesterday, I built a neural network from scratch — just Python and NumPy. No TensorFlow. No PyTorch. Every piece written by hand: forward pass, backpropagation, gradient descent. It recognizes handwritten digits with 96% accuracy.
Here's how.
The Setup
I opened Google Colab on my phone browser, loaded the MNIST dataset (60,000 handwritten digit images), and started coding.
The network has three layers:
- Input: 784 neurons (each pixel of a 28x28 image)
- Hidden: 64 neurons with sigmoid activation
- Output: 10 neurons (one for each digit 0-9)
The Hard Part
Writing backpropagation by hand. When the network makes a wrong prediction, you need to figure out which weights caused the error and adjust them. The math isn't complicated — it's chain rule from calculus — but implementing it without a framework means you understand exactly what's happening.
No model.fit(). No magic. Just matrix multiplication and derivatives.
The Results
After 20 passes through the training data:
- 95.75% accuracy on test data
- Correctly identifies most digits on the first try
- Struggles most with 5s and 8s (they look similar even to humans)
I submitted my predictions to the Kaggle Digit Recognizer competition and got a score of 0.947.
What I Actually Learned
Before this project, I could tell you "neural networks learn from data." After building one from scratch, I can tell you exactly HOW:
- Data goes in, gets multiplied by weights, passes through an activation function
- The output gets compared to the correct answer
- The error flows backward through the network (backpropagation)
- Each weight gets adjusted slightly to reduce the error
- Repeat 60,000 times
That's it. That's deep learning at its core. Frameworks hide this behind one line of code. Building it yourself means you actually get it.
Why I Did This
I'm applying to study AI at university. I wanted my GitHub to show that I understand the fundamentals — not just that I can call an API.
I also built a Telegram bot called ScholarFinder that helps students find fully-funded scholarships. 50+ scholarships searchable by level, field, and region. If you're a student looking for funding, try it.
The Code
Everything is on GitHub: mnist-neural-network
It's one Python file. No dependencies beyond NumPy. If you want to understand how neural networks work under the hood, read it.
What's Next
- Adding ReLU activation to improve accuracy
- Building a second hidden layer
- Trying convolutional neural networks (still from scratch)
- Improving my Kaggle ranking
If I can build this on a phone in Ghana with no laptop, you can build it too. Stop watching tutorials. Open a notebook and start.
Top comments (0)