All projects

Machine learning · Collaborative project

Neural networks, from scratch.

Understanding the machinery behind a prediction.

CONTEXTCollaborative academic project · IF3270
BUILT WITHPython / NumPy / scikit-learn / Jupyter
01 / Conceptual illustration of the project. Not a product screenshot or measured result.

OVERVIEW

Building the model without a high-level training library makes each part of learning inspectable. This project implements the network’s core numerical operations in Python and NumPy, then uses controlled experiments to explore how configuration affects classification.

Build the forward and backward passes

Dense layers support configurable widths and depths. Linear, ReLU, sigmoid, tanh, and softmax activations pair with MSE, binary cross-entropy, or categorical cross-entropy losses.

Backpropagation applies the chain rule to compute gradients. Gradient descent and Adam update the weights; L1 and L2 regularization make the regularization choices explicit. The project also includes a computational-graph autograd implementation.

Configuration is part of the experiment

The implementation includes zero, uniform, and normal initialization, plus Xavier and He options. Leaky ReLU and ELU extend the activation choices. These options make it possible to compare training behavior without changing the core model.

The network drawing illustrates a configurable dense architecture. Its node counts are illustrative, not a claim about a particular trained model or saved experiment.

Compare behavior, not just a final score

The experiment notebook investigates depth, width, activations, learning rates, regularization, and Adam versus basic optimization using the Global Student Placement & Salary dataset.

scikit-learn is used for preprocessing and as a comparison through MLPClassifier, rather than as the custom model’s training engine. No accuracy or loss numbers are reproduced here without the corresponding experiment context. RMSNorm is not implemented.

SOURCES & CONTEXT

Verified against the public README and implementation files. The network illustration is conceptual and does not show benchmark results. Source link points to the owner’s collaborative fork; its README identifies the upstream project.

Inspect project source
NEXT PROJECT

Little Alchemy Solver