This project implements a supervised learning system designed to classify handwritten digits (0-9) from the Scikit-learn digits dataset. By mimicking basic neural architectures, the system learns to interpret low-resolution (8x8 pixel) grids—simulating the type of pattern recognition required for noisy sensor data in aerospace systems.
The dataset consists of 1,797 samples of 8x8 grayscale images. At this resolution, digits are often ambiguous. This project demonstrates how a Multi-Layer Perceptron (MLP) can achieve high-fidelity classification despite low-fidelity input.
Above: A sample of the 8x8 pixel training data showing the "noisy" nature of the input.
- Model: Multi-Layer Perceptron (ANN).
- Optimization: Stochastic Gradient Descent (SGD) with Sigmoid activation.
To find the most efficient architecture, I ran an optimization loop testing 30 different hidden layer configurations (ranging from 5 to 34 neurons).
| Metric | Baseline Model | Optimized Model | Improvement |
|---|---|---|---|
| Hidden Neurons | 5 | 34 | +580% Complexity |
| Accuracy | ~78.4% | 92.6% | +14.2% Gain |
To understand the mathematical "separability" of the data, I compared four different techniques:
- PCA (Linear): Capturing maximum variance.
- t-SNE (Probabilistic): High-fidelity clustering of similar digits.
- Isomap (Manifold): Preserving the "geodesic" distance along curved data structures.
- Spectral Embedding (Graph Theory): Mapping data based on connectivity.
To understand why the Neural Network was successful, I utilized four unsupervised learning methods to project the 64-dimensional pixel data onto a 2D plane. This proves the mathematical "separability" of the digit clusters.
| Method | Visualization | Logic |
|---|---|---|
| t-SNE |  |
Highest Fidelity: Created distinct "neighborhoods" for each digit. |
| PCA |  |
Linear Variance: Shows the primary axes of data spread. |
| Isomap |  |
Manifold Learning: Preserves the curved structure of the data. |
| Spectral |  |
Graph Theory: Maps connectivity between similar pixel patterns. |
- Complexity vs. Accuracy: Increasing the hidden layer size from 5 to 34 neurons showed a clear trend in stabilizing model performance.
- Clustering: t-SNE provided the most distinct separation, suggesting that digit recognition is highly dependent on local neighborhood relationships in the pixel space.
- Language: Python 3.x
- Libraries: NumPy, Matplotlib, Scikit-learn
- Editor: Visual Studio Code (Jupyter Extension)