--- license: mit tags: - image-classification - mnist - numpy - from-scratch - educational datasets: - mnist metrics: - accuracy --- # Four-Layer Neural Network from Scratch (NumPy) A fully-connected network built with **only NumPy**: hand-derived backpropagation and mini-batch SGD, no deep-learning framework. It reaches **97.83% test accuracy** on MNIST and trains in about 88 seconds on a Colab CPU. ## Model Details - **Author:** Vishwajeet Adkine - **Architecture:** `784 → 128 → 64 → 32 → 10` (four weight layers, 111,146 parameters) - **Activations / loss:** ReLU hidden layers, softmax output, cross-entropy with optional L2 - **Initialisation:** He (normal, variance 2 / fan-in) - **Implementation:** NumPy `NeuralNetwork` class with `forward`, `backward`, `step` and `fit`; backprop verified against numerical gradients before training ## Intended Use Learning and demonstration: seeing exactly how forward propagation, backpropagation and SGD work, without framework abstractions. **Not intended** for production use. ## Training | | | |---|---| | Data | MNIST (90% train / 10% validation split of the training set, official test set held out) | | Preprocessing | pixels scaled to [0, 1], flattened to 784 | | Optimiser | mini-batch SGD, batch size 128 | | Learning rate | 0.1 with ×0.95 decay per epoch | | Regularisation | L2, λ = 1e-4 | | Epochs | 30 (default notebook config) | | Hardware | Colab CPU | ## Results | Metric | Value | |--------|-------| | Test accuracy | **97.83%** | | Test loss | 0.0824 | | Training time | 88.2 s (Colab CPU) | ## Limitations - Fully-connected layers ignore image structure, so accuracy is below convolutional models (the CIFAR-10 CNN project in this series covers that step). - Plain SGD with a fixed decay: no momentum, Adam, dropout or batch normalisation. - Evaluated on MNIST only, which is an easy, clean benchmark. Do not expect this accuracy on other digit images (different handwriting, backgrounds or scanning). - No robustness or calibration analysis. ## How to Use ```python net = load_weights("nn_weights.npz") # helper defined in the notebook pred = net.predict(X_test[:5]) ``` ## Licence Code and weights: MIT. MNIST is by LeCun, Cortes and Burges.