LeNet-5, a convolution neural network (CNN), for digit recognition was replicated in PyTorch and trained on the MNIST dataset on Jupyter Notebook with feature visualisation via a gradient-based method. The rectified linear unit (ReLU) was used instead of the hyperbolic tangent function mentioned in the orginal paper. An accuracy of over 97% on the test dataset was achieved by model "LeNet-JFW4E".
- Hyperparameter tuning via Bayesian optimization.
- Implementation of the segmenter to allow for the extraction of multiple digits from noisy images.