{2025} CS 7643 Quiz 1,2 & 3 Questions and Answers Latest 2025 Guide Georgia Institute of T, Exams of Nursing

Download {2025} CS 7643 Quiz 1,2 & 3 Questions and Answers Latest 2025 Guide Georgia Institute of T and more Exams Nursing in PDF only on Docsity!

Guide Georgia Institute of Technology graded A 2025 LATEST

VERSION

Guide Georgia Institute of Technology graded A 2025 LATEST

VERSION

Guide Georgia Institute of Technology graded A 2025 LATEST

VERSION

Guide Georgia Institute of Technology graded A 2025 LATEST

VERSION

CS 7643: Quiz 3 Study online at https://quizlet.com/_dt689k

What is the equa tion for the num ber of parame ters in a convolu tional layer? (kernel_height * kernel_width * input_channels + 1) * num_filters

How do you find the gradient of the weights and input for a convo lutional layer? To compute dw:

1. For the upstream gradient dout, multiply each error d by it's corresponding "kernel overlay" in the original image
2. Add all these together. This should be the same shape as the kernel To compute dx:
3. For the upstream gradient dout, multiply each error d by the kernel
4. Map each receptive field back into the image, and sum overlap

Guide Georgia Institute of Technology graded A 2025 LATEST

VERSION

sponding change in the output. In CNNs, the most common example is pooling layers. Shifting objects in an image slightly produces no change in activation, resulting in translation invariance. Equivariance refers to a change in the input producing some equivalent change 1 / 3 CS 7643: Quiz 3 Study online at https://quizlet.com/_dt689k in the output. The most common example is convolution layers. For example, if an edge detected in the top right corner of one image is present in the top left corner of another, the corresponding activation should shift to the left.

What are saliency maps? Saliency mapping is a technique where the gradient is computed with respect to the original image. Pixels with high absolute gradient values are then colored, which allow you to visualize regions of input pixels that are considered "important".

What is guid ed backpropaga tion? A technique which modifies standard backpropagation by only propagating posi tive gradients for the rectified units. Results in cleaner and sharper visualizations. Helps understand which parts of the image contribute positively to the model's classification.

Guide Georgia Institute of Technology graded A 2025 LATEST

VERSION

10. What is a decon volution? Starting from a feature map to a higher layer, it maps it back to the pixel space. It uses transposed versions of the filters that were applied in the original convolution layers. It allows you to visualize what features an individual filter is capturing.
11. What is Deep Dream? How does it work? Starts with an image and iteratively modifies it to maximize activations. The process is guided by backpropagation. It produces psychedelic images that reveal patterns that the network likes to see, and can reveal what features the network has learned.

2 / 3 CS 7643: Quiz 3 Study online at https://quizlet.com/_dt689k What are Class Activation Map pings? Relies on the global average pooling layer in a CNN. It weights the final convolution layer's feature maps by the weights of the output layer to create a coarse heat map of the important regions. Shows discriminative image regions used by a CNN to identify specific classes.

13. What is Grad-CAM?

Guide Georgia Institute of Technology graded A 2025 LATEST

VERSION

"kernel overlay" in the original image

2. Add all these together. This should be the same shape as the kernel To compute dx:
3. For the upstream gradient dout, multiply each error d by the kernel
4. Map each receptive field back into the image, and sum overlap

Summarize AlexNet Used 5 convolutional layers and 3 fully connected layers. Was the first to use ReLU activations. Started with large filters and then narrowed as network grew deeper.

Summarize VG GNet 16 or 19 layers, often stacked into "blocks" with max pooling. Emphasized the use of multiple 3x3 filters

Summarize In ception Net Uses inception modules, which compute multiple filter sizes at each layer

Summarize ResNet Uses residual blocks, which contain "skip connections", allowing gradients to flow better in deeper networks. Can be very deep, up to 100+ layers

Guide Georgia Institute of Technology graded A 2025 LATEST

VERSION

Explain invari ance and equi variance Invariance refers to the property where a change in the input produces no corre sponding change in the output. In CNNs, the most common example is pooling layers. Shifting objects in an image slightly produces no change in activation, resulting in translation invariance. Equivariance refers to a change in the input producing some equivalent change 1 / 3 CS 7643: Quiz 3 Study online at https://quizlet.com/_dt689k in the output. The most common example is convolution layers. For example, if an edge detected in the top right corner of one image is present in the top left corner of another, the corresponding activation should shift to the left.

What are saliency maps? Saliency mapping is a technique where the gradient is computed with respect to the original image. Pixels with high absolute gradient values are then colored, which allow you to visualize regions of input pixels that are considered "important".

What is guid ed backpropaga

Guide Georgia Institute of Technology graded A 2025 LATEST

VERSION

layer's feature maps by the weights of the output layer to create a coarse heat map of the important regions. Shows discriminative image regions used by a CNN to identify specific classes.

13. What is Grad-CAM? An improvement on Class Activation Mappings that instead uses gradients flowing into the last convolution layer of the CNN to produce a heat map. 3 / 3 CS 7643: Quiz 3 Study online at https://quizlet.com/_dt689k

What is the equa tion for the num ber of parame ters in a convolu tional layer? (kernel_height * kernel_width * input_channels + 1) * num_filters

How do you find the gradient of the weights and

Guide Georgia Institute of Technology graded A 2025 LATEST

VERSION

input for a convo lutional layer? To compute dw:

1. For the upstream gradient dout, multiply each error d by it's corresponding "kernel overlay" in the original image
2. Add all these together. This should be the same shape as the kernel To compute dx:
3. For the upstream gradient dout, multiply each error d by the kernel
4. Map each receptive field back into the image, and sum overlap

Summarize AlexNet Used 5 convolutional layers and 3 fully connected layers. Was the first to use ReLU activations. Started with large filters and then narrowed as network grew deeper.

Summarize VG GNet 16 or 19 layers, often stacked into "blocks" with max pooling. Emphasized the use of multiple 3x3 filters

Summarize In ception Net Uses inception modules, which compute multiple filter sizes at each layer

Guide Georgia Institute of Technology graded A 2025 LATEST

VERSION

which allow you to visualize regions of input pixels that are considered "important".

What is guid ed backpropaga tion? A technique which modifies standard backpropagation by only propagating posi tive gradients for the rectified units. Results in cleaner and sharper visualizations. Helps understand which parts of the image contribute positively to the model's classification.

10. What is a decon volution? Starting from a feature map to a higher layer, it maps it back to the pixel space. It uses transposed versions of the filters that were applied in the original convolution layers. It allows you to visualize what features an individual filter is capturing.
11. What is Deep Dream? How does it work? Starts with an image and iteratively modifies it to maximize activations. The process is guided by backpropagation. It produces psychedelic images that reveal patterns that the network likes to see, and can reveal what features the network has learned.

2 / 3 CS 7643: Quiz 3 Study online at https://quizlet.com/_dt689k

Guide Georgia Institute of Technology graded A 2025 LATEST

VERSION

What are Class Activation Map pings? Relies on the global average pooling layer in a CNN. It weights the final convolution layer's feature maps by the weights of the output layer to create a coarse heat map of the important regions. Shows discriminative image regions used by a CNN to identify specific classes.

13. What is Grad-CAM? An improvement on Class Activation Mappings that instead uses gradients flowing into the last convolution layer of the CNN to produce a heat map. 3 / 3 CS 7643: Quiz 3 Study online at https://quizlet.com/_dt689k

What is the equa tion for the num ber of parame ters in a convolu tional layer?

Guide Georgia Institute of Technology graded A 2025 LATEST

VERSION

Summarize In ception Net Uses inception modules, which compute multiple filter sizes at each layer

Summarize ResNet Uses residual blocks, which contain "skip connections", allowing gradients to flow better in deeper networks. Can be very deep, up to 100+ layers

Explain invari ance and equi variance Invariance refers to the property where a change in the input produces no corre sponding change in the output. In CNNs, the most common example is pooling layers. Shifting objects in an image slightly produces no change in activation, resulting in translation invariance. Equivariance refers to a change in the input producing some equivalent change 1 / 3 CS 7643: Quiz 3 Study online at https://quizlet.com/_dt689k in the output. The most common example is convolution layers. For example, if an edge detected in the top right corner of one image is present in the top left corner of another, the corresponding activation should shift to the left.

Guide Georgia Institute of Technology graded A 2025 LATEST

VERSION

What are saliency maps? Saliency mapping is a technique where the gradient is computed with respect to the original image. Pixels with high absolute gradient values are then colored, which allow you to visualize regions of input pixels that are considered "important".

What is guid ed backpropaga tion? A technique which modifies standard backpropagation by only propagating posi tive gradients for the rectified units. Results in cleaner and sharper visualizations. Helps understand which parts of the image contribute positively to the model's classification.

10. What is a decon volution? Starting from a feature map to a higher layer, it maps it back to the pixel space. It uses transposed versions of the filters that were applied in the original convolution layers. It allows you to visualize what features an individual filter is capturing.
11. What is Deep Dream? How does it work? Starts with an image and iteratively modifies it to maximize activations. The process is guided by backpropagation. It produces psychedelic images that reveal patterns