This repository features a from-scratch Convolutional Neural Network (CNN) implementation. By avoiding high-level frameworks like PyTorch or TensorFlow, I engineered a custom backpropagation engine to master the fundamental calculus and matrix operations governing Deep Learning.
The model was rigorously tested on standard handwritten digits and complex alphanumeric sets.
| Dataset | Classes | Training Acc | Testing Acc |
|---|---|---|---|
| MNIST | 10 | 99.98% | 98.88% |
| EMNIST | 47 | 92.16% | 86.04% |
The architecture is purposefully lean, designed to maximize generalization through a single-pass convolutional feature extractor:
- Conv Layer: 16 Learned Kernels (3x3), ReLU Activation, Max Pooling (2x2).
- Dense ANN: Flatten -> 128 Hidden -> 64 Hidden -> Output Layer.
To achieve production-grade performance using pure NumPy, I implemented several advanced computational strategies:
⚡ Numba JIT Compilation: Wrapped core matrix operations in
@njit(fastmath=True)to execute at C-level speeds, overcoming the Python interpreter bottleneck.🧬 Synthetic Data Augmentation: Expanded the feature space by applying affine transformations via SciPy to 25k samples.
🛡️ L2 Regularization & Decay: Stabilized gradient descent and closed a 10% overfitting gap by manually deriving and applying L2 penalties to the weight updates:
(Where
To truly understand what the CNN is "learning," we have to look inside the hidden layers.
In this architecture, I implemented 16 custom kernels (3x3 matrices). During the forward pass, these kernels convolve (slide) across the 28x28 input image. Each kernel acts as a specialized filter, trained strictly via backpropagation to detect specific low-level features such as horizontal lines, vertical edges, or sharp loops.
When an image is passed through these 16 kernels and the ReLU activation function (which drops negative values to introduce non-linearity), it produces 16 distinct "Feature Maps."
Below are visualizations of what the CNN actually "sees" inside its first layer when looking at a handwritten digit. Notice how the bright yellow areas indicate high activation—this is where a specific kernel successfully found the geometric shape it was trained to look for!
To run this model locally and see the training in action:
# 1. Clone the repository
git clone [https://github.com/Jit338/E-MNIST.git](https://github.com/Jit338/E-MNIST.git)
# 2. Navigate into the directory
cd E-MNIST
# 3. Install required mathematical packages
pip install numpy numba scipy scikit-learn matplotlib
# 4. Launch the Jupyter Notebook
jupyter notebook CNN2.ipynb