Final Flashcards
If we build a neural network with no activation function, we get a nonlinear classifier. (T or F)
False
What does a neuron compute?
A neuron computes a linear function (z = Wx +b) followed by an
activation function.
During the backwards pass, each node in the graph receives
_______ gradients and multiplies them by __________ gradients to
compute _________ gradients.
Upstream, local, downstream
What is the benefit of using a Pooling Layer in a CNN?
Reduces the dimensionality of each feature map
The convolution operation shrinks the matrix of pixels (input image)
only if the size of the filter is greater than 1.
True
In _______ layer, every neuron in the previous layer is connected to
each and every neuron in the next layer.
Fully connected
Generative Adversarial Networks (GANs) jointly train generator and ____________ with a minimax game.
c. Discriminator
Match each of the following models with its corresponding:
- Assign labels to data
- Feature Learning (with labels)
Discriminative Model: Learn a probability distribution p(y|x)
Match each of the following models with its corresponding:
- Assign labels, while rejecting outliers
- Generate new data conditioned on input labels
Conditional Generative Model: Learn p(x|y)
Match each of the following models with its corresponding:
- Detect outliers
- Feature learning (without labels)
- Sample to generate new data
Generative Model: Learn a probability distribution p(x)
What is the problem described below?
Do well on training data, but poorly on validation data due to variance.
Overfitting
What is the given definition below?
Label each pixel in the image with a category label. It only cares about pixels.
Semantic Segmentation
We developed a segmentation algorithm, and we like to check how good our segmentation prediction is. How can we compare our segmentation prediction to the ground-truth?
Intersection over Union
When we visualize the first layer of a deep network, what will we see in the output of these layers?
Low level features such as edges and corners
Which visualization technique is used to understand the flow of gradients through different layers of a Convolutional Neural Network (CNN)?
Backpropagation
Which of the following techniques is used to highlight the most relevant regions in an inpt image for explaining the predictions of a Convolutional Neural Networks (CNNs)?
Grad-CAM
List the SIFT Descriptors for feature detection in order:
1) Normalize rotations
1) Normalize the rotation / scale of the patch
2) Compute gradient at each pixel
3) Divide into sub-patches (here 2x2, actually 4x4)
4) In each sub-patch, compute histogram of 8 gradient directions
5) Describe the patch with 448 = 128 numbers
What Components of a Convolutional Network is this?
Full Connected Layer
What Components of a Convolutional Network is this?
Activation Function
What Components of a Convolutional Network is this?
Convolutional Layers
What Components of a Convolutional Network is this?
Pooling Layer
What Components of a Convolutional Network is this?
Normalization
List the steps for Training a Convolutional Network and what does each step do?
1) Download big datasets: acquire and download large dataset
2) Design CNN architecture: Make a CNN from scratch or modify existing CNN’s like Alexnet, Vgg, or googlenet
3) Initialize Weights: set weights to desired initial values
4) For t = 1 to T:
1. Form minibatch: form a small
batches of data from the training
set used to trained the model.
2. compute loss + gradient: The model's prediction 3. Update Weights
What Is Morphology?
Morphological image processing (or
morphology) describes a range of image
processing techniques that deal with the shape
(or morphology) of features in an image
Morphological operations
typically applied
to remove imperfections introduced during
segmentation, and so typically operate on bi-
level images
17
two basic morphological
erosion and dilation
More interesting morphological operations can
be performing combinations of
erosions and dilations
The most widely used of these compound
operations are:
– Opening
– Closing
- Roberts filter
- Prewitt filter
- Sobel filter
Name a voting scheme
Hough transform
In backward propagation what happens in the backward pass?
During the backward pass, each node in the graph receives upstream gradients and multiplies them by local gradients to compute downstream gradients
used for Circle Detection
Hough transforms
region:
no change in
all directions
“flat
no change along
the edge
direction
edge
significant
change in all
directions
corner
SIFT Descriptors
- Normalize the rotation / scale of the patch
- Compute gradient at each pixel
- Divide into sub-patches (here 2x2, actually 4x4)
- In each sub-patch, compute histogram of 8
gradient directions - Describe the patch with 448 = 128 numbers
What happens in backpropagation during backward pass?
During the backward pass, each node in the graph receives upstream gradients and multiplies them by local gradients to compute downstream gradients
Fill in the blanks
What are the different machine vision tasks?
- Classification - no spatial extent
- semantic segmentation - no objects, just pixels
- object detection:
What is a Discriminative Model
Assign labels to data, Feature Learning with labels
What is a Generative Model
Detect outliers, Feature Learning without labels
What is a Conditional Generative Model
Assign labels, while rejecting outliers
What is GANs?
Generative Adversarial Networks
What does GANs do?
Jointly train generator and discriminator with minimax game