Unit 1: Foundations of Computer Vision and Deep Learning - Practice Quiz

1 Which task is a common application of computer vision?

2 What is the main goal of computer vision?

3 How are visual features commonly obtained in traditional computer vision?

4 What is a key advantage of deep-learning-based vision systems?

5 How is a digital image commonly represented inside a computer?

6 What is a pixel?

7 How many color channels does a standard RGB image have?

8 What do the letters in the RGB color space represent?

9 Which color space represents an image using hue, saturation, and value?

10 What does resizing an image change?

11 What happens when an image is cropped?

12 What is a common result of dividing 8-bit pixel values by ?

13 What do neural network weights represent?

14 During neural network training, what is learned from the data?

15 What does a basic perceptron compute before applying its activation function?

16 In a feedforward neural network, which direction does information move?

17 What is the output of the ReLU function when its input is negative?

18 Why are activation functions used in neural networks?

19 What does a loss function measure during training?

20 What is the purpose of gradient descent in neural network training?

21 A warehouse robot must identify damaged packages and determine their locations on a conveyor belt. Which computer vision task is most directly suited to this requirement?

22 A traditional image classifier uses manually designed edge and texture descriptors, while a convolutional neural network is trained on raw images. What is the main advantage of the deep-learning approach?

23 An RGB image has dimensions and uses 8 bits for each channel value. How much uncompressed storage does it require?

24 A grayscale image uses pixel values from 0 to 255. If a pixel changes from 40 to 200, what is the most direct interpretation?

25 A model expects input tensors in the shape . How should an RGB image stored as be rearranged?

26 A vision system must detect brightly colored road signs under varying illumination. Which color space is useful for separating color type from brightness?

27 A image is resized directly to without preserving its aspect ratio. What is the most likely effect?

28 During training, random crops are taken from large images before classification. What is a likely benefit of this procedure?

29 Pixel intensities are stored as integers from 0 to 255. Which transformation maps them linearly to the interval ?

30 A dense layer receives 64 input features and produces 10 outputs, with one bias per output. How many trainable parameters does it contain?

31 A perceptron computes and outputs 1 when , otherwise 0. What is its output for and ?

32 Why can a feedforward network with nonlinear hidden layers represent decision boundaries that a single perceptron cannot?

33 For the ReLU activation , what gradient is normally used during backpropagation when ?

34 The output layer of a network must produce probabilities for five mutually exclusive classes. Which activation is most appropriate?

35 A three-class classifier predicts probabilities , and the true class is the second class. What is the categorical cross-entropy loss for this sample?

36 A model predicts continuous bounding-box coordinates. Which loss is commonly suitable for measuring basic coordinate regression error?

37 A parameter has value , its gradient is , and gradient descent uses a learning rate of . What is the updated parameter value?

38 Training loss oscillates strongly and sometimes increases after parameter updates. Which adjustment is the most reasonable first response?

39 A small medical-image dataset contains only a few hundred labeled samples. Which strategy best uses a deep-learning model while reducing data requirements?

40 A pretrained model was trained using per-channel mean and standard deviation normalization. What is the likely consequence of supplying only raw 0-to-255 pixel values during inference?

41 A calibrated disease-screening model is deployed in a population where disease prevalence is much lower than in the validation population. Assume the class-conditional score distributions remain unchanged. Which consequence is most likely if the original decision threshold is retained?

42 A manufacturer must classify parts from only 80 labeled images. Parts may rotate arbitrarily, and engineers know that edge-orientation distributions are highly diagnostic. Which approach has the strongest inductive bias for this setting?

43 An image stored as an interleaved HWC array is passed to a framework expecting CHW format. The programmer uses reshape(C, H, W) instead of transposing axes. What is the principal result?

44 An RGBA image contains fully transparent pixels whose stored RGB values are bright red. It is resized using bilinear interpolation and later composited over black. Which procedure best prevents red fringes near opaque object boundaries?

45 Ignoring biases, compare a standard convolution with 32 input channels and 64 output channels against a depthwise-separable version with a depthwise stage and a pointwise stage. How many weights does the separable version use?

46 A black pixel and a white pixel, represented by sRGB values 0 and 1, are averaged. Approximately what sRGB value represents their physically correct equal-energy mixture, assuming the usual sRGB transfer function?

47 A one-pixel black-and-white checkerboard is downsampled by a factor of 8. Which operation most directly prevents aliasing while preserving the correct average intensity?

48 An image of size is cropped using the rectangle with top-left corner and size . A bounding box in the original image is in format. After clipping to the crop and converting to crop-local coordinates, what is the box?

49 A pretrained network expects channel normalization , but deployment supplies . Retraining is unavailable. Which modification can exactly preserve the network's output, assuming normalization is channelwise affine and the first layer is affine?

50 Every neuron in a hidden layer is initialized with the same weight vector and bias. Under deterministic full-batch gradient descent, what typically happens to those neurons?

51 A single perceptron with a threshold activation must classify the four binary inputs so that and are positive, while and are negative. Which statement is correct?

52 Consider with no activation between the two affine layers. If the hidden width is , which statement best characterizes the resulting mapping?

53 For a leaky ReLU with negative slope , what weight variance approximately preserves activation variance in a wide layer whose zero-mean inputs are symmetric, under the usual independence assumptions?

54 For -class cross-entropy with label smoothing target , what probabilities minimize the expected loss for a single deterministic training label, assuming the model can represent any distribution?

55 Why is adding an penalty to the loss generally not equivalent to decoupled weight decay when using Adam?

56 A vision pipeline converts an RGB image to YCbCr 4:2:0 before storage. Which spatial information is reduced by this representation?

57 An image of size is letterboxed into a input while preserving aspect ratio and centering the result. Under the scale-and-pad convention, how should a point be transformed?

58 A CNN with batch normalization is trained correctly, but during inference each batch contains only one highly atypical image and the layer is accidentally left in training mode. What failure is most directly expected?

59 For binary classification with logit , which expression computes binary cross-entropy stably without explicitly evaluating or ?

60 Four microbatches are used to emulate one large batch. Their gradients are summed, and the optimizer performs one SGD update with momentum. Under what adjustment does this match using the mean gradient of the combined large batch, assuming no batch-dependent layers?