Fashion MNIST Classification

An exploratory analysis of Convolutional Neural Networks (CNN) for classification of clothing items from the fashion variant of the famous MNIST dataset. A diverse set of techniques were employed in hyperparameter tuning, a few examples are:

• Early stopping
• GridSearchCV
• Learning decay schedulers
• Dropout layers
• Optimizers and variants
• Kernel initializations
• Altering stride, size, # of filters
• Varying depth and width of CNN

Best Performance Achieved

	Training	Validation	Testing
Accuracy	0.980	0.924	0.924
Loss	0.0596	0.223	0.243

Optimal Architecture

• 1 Convolution Layer: 32 filters of 5x5, stride 1
• 1 Max Pooling Layer: 3x3 pool, stride 1
• 1 Convolution Layer: 32 filters of 5x5, stride 1
• 1 Max Pooling Layer: 2x2 pool, stride 2
• 1 Dense Layer: 1000 nodes, activation = ReLU
• 1 Dropout Layer: 10% (0.1)
• 1 Output Layer: 5 nodes, activation = Softmax

It contains the following hyperparameters:

• Loss: categorical crossentropy
• Optimizer: Adam
• Learning rate: step decay ($LR_i$ = 0.001, $DR$ = 0.6, $E_D$ = 9.0)
• Kernel Initializer: he_uniform
• Batch size: 64
• Epochs: 75