NumPyNet
An ongoing learning project building neural networks from scratch with Python and NumPy. Exploring manual backpropagation, numerical gradient checks, and training behavior through XOR and synthetic multiclass experiments.
Synthetic XOR / spiral data
Dense layers & activations
Manual backpropagation & SGD
Gradient checks & training diagnostics
From an idea to an implementation.
An ongoing learning project building neural networks from scratch with Python and NumPy. Exploring manual backpropagation, numerical gradient checks, and training behavior through XOR and synthetic multiclass experiments.
I am still learning neural-network fundamentals. This personal project helps me work through forward passes, gradients, and optimization step by step, with reproducible experiments and mathematical notes in the repository.
Under the surface.
An educational NumPy implementation with explicit gradients, numerical gradient verification, and experiments comparing learning rates and training-stability techniques. XOR uses four training samples; the multiclass example uses synthetic spiral data with a separate validation set. These experiments support learning and do not establish real-world performance.
Architecture view / inspect components
- Component 1Synthetic XOR / spiral data
- Component 2Dense layers & activations
- Component 3Manual backpropagation & SGD
- Component 4Gradient checks & training diagnostics
This view organizes the known components. Consult the repository for exact interfaces, data shapes and implementation.
Source first. Claims second.
The repository includes numerical gradient checks, automated tests, and reproducible training experiments. XOR results concern only four training samples, and the multiclass demonstration uses synthetic spiral data. I am continuing to learn; these experiments do not establish performance on real datasets. The source repository is available for technical review.
Explore source code