CIFAR-10 Image Classification 😓

• Used Keras
• CNN for classification
• 50000 images for training and 10000 images for testing
• Accuracy with baseline VGG model
  • 1 VGG block - 67%
  • 2 VGG blocks - 71.5%
  • 3 VGG block - 73%

Model:

• VGG model is easy to understand and implement architecture
• Stacking convolutional layers with small 3x3 filters followed by max pooling layer. These form a block which can be repeated and the filters are increasing with depth to the network
• Model is optimized using stochiastic gradient descent
• Used 3 VGG block
• Model overfits test dataset within first 15-20 epochs

Improvements:

• Accuracy increased slightly with increase in VGG blocks
• Will attempt regularization techniques to improve model
  • Added Dropout layers after each max pooling layer with a rate of 20%
    • Accuracy of 82.4%
    • Model converges well for first 40 epochs - no further improvements after. Can add early stopping to save the model on the test set during training when no further improvements are made
    • Increasing dropout with depth in model led to an accuracy of 82.8%
  • Using weight decay/weight regularization did NOT improve accuracy
  • Data augmentation - making small random modifications to copies of examples in training dataset
    • Accuracy of 84.3%
    • Horizontal flips, minor shifts, and small zooming/cropping

[Work in progress]
Using dataset from CIFAR-10

Current Accuracy: 84.3%