Quiz 3: CS 7643 Deep Learning | Questions with Verified Answers | 100% Correct| Latest 2025/2026 Update - Georgia Institute of Technology.

Quiz 3: CS 7643 Deep Learning | Questions
with Verified Answers | 100% Correct| Latest
2025/2026 Update - Georgia Institute of
Technology.

CAM = Class Activation Mapping
i,- i,- use Global Average Pooling i,- i,- i,-i,- i,- i,- i,- i,- i,-



layer as final layer to average the activations of each feature map
i,- i,- i,- i,- i,- i,- i,- i,- i,- i,- i,- i,-



and run through softmax loss layer to highlight the important
i,- i,- i,- i,- i,- i,- i,- i,- i,- i,-



regions of the image by projecting back the weights of the
i,- i,- i,- i,- i,- i,- i,- i,- i,- i,- i,-



output on the convolutional feature maps
i,- i,- i,- i,- i,-




Grad-CAM more versatile version of CAM that can produce
i,-i,- i,- i,- i,- i,- i,- i,- i,- i,- i,-



visual explanations for any arbitrary CNN, even if the network
i,- i,- i,- i,- i,- i,- i,- i,- i,- i,-



contains a stack of fully connected layers tooi,- i,- i,- i,- i,- i,- i,-




let the gradients of any target concept score flow into the final
i,- i,- i,- i,- i,- i,- i,- i,- i,- i,- i,- i,-



convolutional layer; then compute an importance score based on i,- i,- i,- i,- i,- i,- i,- i,- i,-



the gradients and produce a coarse localization map highlighting
i,- i,- i,- i,- i,- i,- i,- i,- i,-



the important regions in the image for predicting that concept
i,- i,- i,- i,- i,- i,- i,- i,- i,-




What regions of image is model looking at to make prediction?
i,- i,- i,- i,- i,- i,- i,- i,- i,- i,-




Which individual regions have highest class activation as you
i,- i,- i,- i,- i,- i,- i,- i,- i,-



extract layer from CNN? Direction/magnitude of gradients to
i,- i,- i,- i,- i,- i,- i,- i,-



determine which gradients are causing the most updates to the
i,- i,- i,- i,- i,- i,- i,- i,- i,- i,-



NN
Objective: inspective given layer of CNN and correlate to output
i,- i,- i,- i,- i,- i,- i,- i,- i,-

, Task specific (if asked what is a dog -> dog pixels are more
i,- i,- i,- i,- i,- i,- i,- i,- i,- i,- i,- i,- i,-



important)


Adversarial examples Inputs formed by applying small but
i,- i,-i,- i,- i,- i,- i,- i,- i,- i,-



intentionally worst-case perturbations to examples from the
i,- i,- i,- i,- i,- i,- i,-



dataset, such that the perturbed input results in the model
i,- i,- i,- i,- i,- i,- i,- i,- i,- i,-



outputting an incorrect answer with high confidence.
i,- i,- i,- i,- i,- i,-




Guided Backprop i,- i,-i,- i,- Layer by layer (deconvolution is similar to
i,- i,- i,- i,- i,- i,- i,-



backprop)
From details to more abstracted representations
i,- i,- i,- i,- i,-




L1 Loss
i,- i,-i,- i,- Sum of Absolute Value of (true - predicted)
i,- i,- i,- i,- i,- i,- i,-




L2 Loss
i,- i,-i,- i,- Sum of Absolute Value of (true - predicted)^2
i,- i,- i,- i,- i,- i,- i,-




Mean Squared Error (MSE)
i,- i,- i,- i,-i,- i,- Average of (true - predicted)^2 i,- i,- i,- i,-




Modeling Error Given a particular NN architecture, the actual
i,- i,-i,- i,- i,- i,- i,- i,- i,- i,- i,-



model that represents the real world may not be in that space.
i,- i,- i,- i,- i,- i,- i,- i,- i,- i,- i,-