Model collapse is the term given to failure of a model to correctly generate or discriminate across a real world distribution when it is repeatedly fed its own generations as training data. Given the ...

The optimized CNN model consists of 4 convolutional layers followed by 1 fully connected layer: Input (1, 28, 28) ↓ Conv2d(1→8, kernel=3) + ReLU + MaxPool2d(2) ↓ Conv2d(8→16, kernel=3) + ReLU + ...