CNN

25 problems

• Conv2D forward (naive, stride 1, no pad)Medium
• Conv2D forward (padding + stride)Medium
• Max pooling 2D forwardEasy
• Average pooling 2D forwardEasy
• Conv2D backward (dx, dW)Hard
• Mini-CNN forward (capstone)Medium
• The classifier head — FC + softmaxMedium
• Feature volume — a conv layer's output shapeEasy
• Convolution — feature-map sizeEasy
• Convolution — one patch stepEasy
• Why CNNs — conv vs dense parameter countMedium
• Convolution — slide a 1D filter (weight sharing)Medium
• Feature map — bias and ReLUEasy
• Global average poolingEasy
• Edges — the horizontal gradientEasy
• Hierarchy — the receptive field grows with depthEasy
• Sobel edge magnitudeMedium
• Which filter fires? (template detection)Easy
• Dense block with 2D convolutionsMedium
• Gradient paths through L residual blocksEasy
• Pre-LN vs post-LN blockMedium
• The residual Jacobian I + ∂f/∂xEasy
• The residual stream grows without LayerNormMedium
• Group normalizationMedium
• Residual block with shortcutEasy