02 · Computer vision / Case study
Sign Language MNIST from first principles
A NumPy implementation of perceptron and feed-forward networks focused on label quality, error analysis and model comparison.
The problem
Understand the mechanics of neural-network learning by implementing the essential parts without a high-level framework.
My contribution
I implemented a perceptron and feed-forward network in NumPy, including backpropagation, mini-batch gradient descent, softmax and cross-entropy. I compared variants and produced learning curves, confusion matrices and a label-diagnostic report.
A simple model before a larger network
The perceptron provides a baseline. A nonlinear network must justify its complexity through its behaviour on the same data.
Mistakes are material for investigation
Accuracy with the raw test labels is 45.84%. A diagnostic alignment of the label mapping gives 88.62%. The report retains both values; the second does not replace the first.
Make the learning process visible
Validation, early stopping, regularisation and learning curves help distinguish a learning problem from a data or evaluation problem.
Evidence and interpretation
What the result actually says.
Primary accuracy on the original held-out labels: 0.4584. Separate aligned-label diagnostic: 0.8862.
The central result is the distinction between a test result and a diagnosis of its cause. 88.62% is an aligned-label diagnostic; 45.84% remains the primary result on the original held-out file.
The score belongs to the Sign Language MNIST benchmark; it is not the performance of a real-world camera-based sign-language system.
The next useful experiment
Validate label provenance and mapping against an independent dataset. Camera input, variable lighting and moving signs would require a separate study.
Descriptions checked against the public project documentation on 17 September 2026. Scores are reported project results; the experiments were not rerun for this website update.