NumpyNet

This project captures a learning exploration into the fundamentals of deep learning. In this project I build and train neural networks from scratch using Numpy, and I validate my approach by comparing it to PyTorch-trained models. Repo can be found here.

---
Skils involved:
- Deep learning
- Numpy
- Backpropagation
- PyTorch
---

Problem description

Neural networks are function approximators comprised of compositions of linear operators and non-linear activation functions. At their core, neural networks combine many y=wx+b operations to minimize the prediction error of the entire neural network on the training data.

In this project I set out to become familiar with the under-the-hood of multi-layer perceptrons (MLP) by implmented methods for creating, training, and testing neural networks with Numpy alone.

My contributions

• Class definitions used for creating models in object-oriented Python
• Full implementations of feedforward and backpropagation operations
• Evaluation on toy dataset and comparison with PyTorch

Data

A small dataset was generated to debug my implementation. The dataset is relative easy classification task consisting of 5000 2-feature observations belonging to 3 separate classes. The data is displayed below:

Results