basedNN

Yet another collection of neural networks implementations from scratch

built with nothing but grit, tears, and raw Python
(using python lists as tensors 💀)

---

Why?

"Why build neural networks from scratch in 2025?"

Because libraries are for the weak 🗿

but also because i wanted a repo with simple and easily readable code to check when i need to refresh the implementation details of the backprop algorithm

---

Features

• ✅ No external libraries: not even NumPy, we do matrix math like it’s 1998
• ✅ Artisanal backpropagation: 100% gluten free handcrafted gradients
• ✅ Readability: so simple, even your toaster could understand it
• ❌ Efficiency: training MNIST? Estimated completion: 2047

---

Usage

to start training an architecture cd into its directory and from there run the main.py in src folder for example to test the mlp:

$ git clone https://github.com/samas69420/basedNN
$ cd basedNN/mlp
$ python src/main.py   

the hyperparameters like learning rate, network layers etc can be modified in main.py

---

Contributing

1. Fork the repo.
2. Add more inefficient (but easy to read) code.
3. ???.
4. Profit (not really).

---

Ze math

for more information about the notation: https://robotchinwag.com/posts/the-tensor-calculus-you-need-for-deep-learning/

---

This repo is intended for educational purposes only and shouldn't be used for real world applications, why would it be a bad idea?

• 🚨 No vectorization: we loop like it’s a cardio workout
• 🚨 No GPU support: your CPU will hate you
• 🚨 No fancy CPU optimizations: this code is not even multithreaded, with large models or datasets it would be slower than continental drift