CNN Cheat Sheet 2026

The 30 highest-yield CNN facts, distilled from real exam questions. Print it, save it as a PDF, or study it here — free, no sign-up.

1. What problem does data augmentation specifically help address during transfer learning fine-tuning? → Overfitting when the target dataset is small
2. What is 'inverted dropout' and why is it preferred in practice? → Scaling kept activations by 1/(1-p) during training so no scaling is needed at test time
3. In few-shot learning via transfer learning, what enables a CNN to generalize from very few examples? → Rich feature representations learned from large source datasets
4. Which layer type is typically removed when adapting a pretrained CNN for a new classification task? → Fully connected output layer
5. What is the total number of layers (depth) in the original GoogLeNet (Inception v1)? → 22
6. How does GoogLeNet reduce computational cost before applying 3x3 and 5x5 convolutions in the Inception module? → By applying 1x1 convolutions as bottlenecks
7. When both batch normalization and dropout are used in the same network, what ordering is generally recommended? → Conv → Batch Norm → ReLU → Dropout (in fully connected layers)
8. GoogLeNet includes auxiliary classifiers during training. What is their primary purpose? → To combat vanishing gradients by injecting gradient signal at intermediate layers
9. Which pooling variant performs a weighted average where closer elements contribute more, based on their distance from the pooling region center? → Weighted average pooling
10. What is the 'dying ReLU' problem? → Neurons with ReLU permanently output zero because their weights push inputs negative
11. What is the purpose of transposed convolutions (deconvolutions) in segmentation networks? → To learn to upsample feature maps back to a higher spatial resolution
12. In what year did ResNet win the ImageNet Large Scale Visual Recognition Challenge (ILSVRC)? → 2015
13. The Leaky ReLU activation function was designed to address which limitation of standard ReLU? → Zero gradient for negative inputs leading to dead neurons
14. Which of the following regularization techniques involves halting the training process when the model's performance on a validation set stops improving? → Early Stopping
15. What problem does Non-Maximum Suppression (NMS) solve in object detection? → Multiple overlapping predictions for the same object
16. Which popular pretrained CNN architecture introduced depthwise separable convolutions, making it efficient for transfer learning on mobile devices? → MobileNet
17. During CNN training, which data augmentation technique generates new training samples by combining two images and their labels linearly? → Mixup
18. In feature map visualization, what does 'channel-wise mean activation' summarize? → The average activation value across all spatial positions in a feature map
19. What is the effect of applying Gaussian blur augmentation to CNN training images? → It simulates out-of-focus conditions, making the model robust to image blurriness
20. During CNN training, which regularization method randomly drops entire feature maps rather than individual neurons? → DropBlock
21. What is the output range of the Tanh activation function? → [-1, 1]
22. What is the primary reason that ensembling multiple independently trained CNNs improves generalization compared to a single model? → Individual models overfit to different patterns; averaging cancels out their errors
23. What is the depth of the winning ResNet model in the ILSVRC 2015 competition? → 152 layers
24. What effect does L1 regularization have on CNN weight vectors compared to L2? → It produces sparse weights with many zeros
25. What is the role of anchor boxes in SSD and Faster R-CNN? → They serve as reference boxes of predefined scales and aspect ratios for regression
26. A CNN trained on 10,000 images achieves 99% training accuracy but only 62% validation accuracy. Which symptom does this describe? → Overfitting
27. Which PyTorch library is most commonly used for applying data augmentation transforms to image datasets? → torchvision.transforms
28. Global Average Pooling (GAP) is frequently used at the end of a CNN to replace which traditional layer? → Fully connected (dense) layer
29. What is 'knowledge distillation' in the context of training CNNs? → Training a smaller student model to mimic the outputs of a larger teacher model
30. What does AutoAugment do in the context of CNN training? → Searches for an optimal augmentation policy using reinforcement learning