Unit 2: Convolutional Neural Networks and Training Techniques - Practice Quiz

1 What is the main purpose of a convolutional neural network in computer vision?

2 What happens during a convolution operation?

3 What does a filter learn to detect in a convolutional neural network?

4 What is a feature map in a convolutional neural network?

5 Why is zero padding commonly added before convolution?

6 What does the stride of a convolution specify?

7 What does max pooling select from each local region?

8 What is the receptive field of a neuron in a CNN?

9 LeNet was originally designed mainly for which task?

10 Which event made AlexNet especially influential in computer vision?

11 Which type of convolutional filter is used extensively in VGG networks?

12 What is the key architectural idea introduced by ResNet?

13 What is a main feature of an Inception module?

14 What is the main goal of regularization during neural network training?

15 What does batch normalization normalize during training?

16 What does dropout do during neural network training?

17 Which operation is a common image data augmentation method?

18 What does a learning rate scheduler change during training?

19 Why is proper weight initialization important in a neural network?

20 When does early stopping usually end model training?

21 A CNN receives a grayscale image and applies 10 filters of size with stride 1 and no padding. What is the spatial size and depth of the resulting feature maps?

22 A input is convolved with a filter using stride 1 and no padding. How many values are produced in the output feature map?

23 During early CNN training, a filter produces high activations mainly along vertical edges. What does this indicate about the filter?

24 Why can a feature map in a deeper CNN layer represent more complex visual patterns than one in the first layer?

25 A image is processed with a filter using stride 1. What padding size preserves the spatial dimensions of the image?

26 An input feature map has spatial size . A filter is applied with stride 2 and no padding. What is the output spatial size?

27 A feature map is processed using max pooling with stride 2. What is the output size, and what does each output value represent?

28 Two consecutive convolutional layers use filters, stride 1, and no pooling. What is the receptive field size of one activation in the second layer relative to the original input?

29 Which design choice is most characteristic of the original LeNet architecture?

30 Which combination of techniques helped AlexNet achieve strong performance on the ImageNet classification task?

31 What is a key architectural characteristic of VGG networks?

32 A residual block computes . What is the primary benefit of the shortcut connection?

33 Why does an Inception module apply convolutions with different kernel sizes in parallel?

34 A CNN achieves very low training error but substantially higher validation error. Which intervention most directly addresses this problem?

35 During training, batch normalization is inserted between a convolutional layer and its activation. What is its main purpose?

36 During training, dropout randomly sets a fraction of activations to zero. Why can this improve validation performance?

37 Which augmentation is generally suitable for an object-recognition dataset when the object can appear at different horizontal positions but its class is unchanged?

38 A model's validation loss stops improving while the learning rate remains relatively large. Which scheduling strategy may help refine the solution?

39 A CNN uses ReLU activations. Which initialization is generally appropriate for maintaining activation variance across layers?

40 Training loss continues decreasing, but validation loss has increased for several consecutive epochs. What is the most appropriate early-stopping action?

41 A fully convolutional network contains only stride- convolutions, pointwise nonlinearities, and no padding effects. A global average pooling layer is then applied to its final feature maps. Which statement best describes the resulting transformation?

42 An input of size is processed by a grouped convolution with output channels, a kernel, dilation , padding , stride , and groups. Biases are disabled. What are the output shape and number of trainable weights?

43 A convolutional filter and its bias produce one channel followed by ReLU. Which parameter transformation leaves the network function unchanged if the next layer is linear with respect to that channel?

44 At each spatial location, a -channel feature map is transformed by a convolution to channels and then by another convolution back to channels. No nonlinearities or biases are used. What is the maximum rank of the resulting channel-wise linear transformation?

45 A one-dimensional convolution uses input length , kernel size , dilation , and stride . Under the SAME_UPPER convention, what padding is added to obtain output length ?

46 Consider a convolution with stride on an infinite input grid. Ignoring nonlinearities, which translation property is guaranteed for arbitrary convolutional weights?

47 A max-pooling layer has stride and no padding. An interior activation is the unique maximum in every pooling window that contains it. If the upstream gradient from each such output is , what gradient reaches this activation?

48 A network applies, in order: a convolution with stride , a convolution with stride , a convolution with dilation and stride , and a pooling layer with stride . All other dilations are . What are the final receptive-field size and input-space jump?

49 The original LeNet-5 spatial pipeline is applied unchanged to a input: valid convolution, pooling, valid convolution, pooling, and a valid C5 layer. What happens?

50 AlexNet historically used grouped convolutions to distribute computation across two GPUs. In a convolution with two groups, what is the direct connectivity consequence?

51 Assume constant channel width , stride , and no biases. Compared with one convolution, what is achieved by three consecutive convolutions?

52 For a residual block , which initialization most directly makes the block begin as an identity mapping while retaining trainable parameters in the residual branch?

53 An Inception module receives channels and has four branches: with outputs; with outputs followed by with outputs; with outputs followed by with outputs; and pooling followed by with outputs. Ignoring biases, what are the concatenated output depth and total convolutional weight count?

54 Why is adding an penalty to the loss generally not equivalent to decoupled weight decay when training with Adam?

55 During training, a bias-free convolution is immediately followed by batch normalization. Ignoring , what happens if all weights of one output filter are multiplied by a positive constant ?

56 In inverted dropout, an activation is dropped with probability and otherwise divided by the keep probability. What are the expectation and variance of the training-time output, conditional on fixed ?

57 For semantic segmentation with integer-valued class masks, which augmentation pipeline correctly applies a random rotation to each image-mask pair?

58 A cosine schedule without restarts uses . If , , , and , what is ?

59 A ReLU convolution has a kernel and input channels. Under He normal initialization, what variance and approximate standard deviation should each weight have?

60 A model is evaluated on the same validation set after every epoch, and the checkpoint with the lowest validation loss is selected. Which evaluation protocol best estimates the selected model's generalization performance?