Quiz

Question 1

Semantic Segmentation

Answer 1

assigns a class label to each pixel, grouping all objects of the same class (all cars as one entity)

Question 2

Instance segmentation

Answer 2

identifies each object instance separately, even if they belong to the same class (distinguish between diff cars)

Question 3

Fully Convolutional network (dense prediction)

Answer 3

• classify each pixel in the input image into different class labels
• encoder (convolution network)
• decoder (deconvolution network)

Question 4

Unsampling methods in Decoder

Answer 4

• transposed convolution
• backward-stride convolution
• max unpooling

Question 5

U-Net

Answer 5

Fuses upsampled feature maps in the decoder with corresponding feature maps from the encoder to preserve spatial details.

Question 6

Dilated Convolution

Answer 6

uses a large sparse filter to increase receptive field

Question 7

Mask R-CNN for instance segmentation

Answer 7

Faster R-CNN + fully convolutional network for semantic
segmentation

Question 8

Hidden states

Answer 8

• receives inputs from previous layer and hidden state
• weights are same across hidden states in the same hidden layer
• weights are different across different hidden layers

Question 9

Basic RNN’s problems

Answer 9

Exploding gradients

Vanishing gradients

Question 10

LSTM (Long Short-Term Memory)

Answer 10

• Sigmoid functions are used to measure importance of input
• forget and input gates decide the cell state
• output gate produces output of the cell

Question 11

RNN with Attention

Answer 11

• Context vectors are used as the input to each timestep of the decoder
• Attention provides different context vectors by calculating alignment scores between decoder states and features from the encoder

Question 12

Self-attention

Answer 12

Enabled each element in a sequence to interact and learn dependencies with every other element, irrespective of their distance

Question 13

Vision transformer

Answer 13

• Split an image into patches and represent patches as lower-dimensional embeddings and add positional embedding
• apply self-attention
• replace convolutional layers
• trained on a large dataset

Question 14

Semi-supervised learning

Answer 14

• partially labeled data in the training dataset
• get pseudo label by applying the trained model using labeled data to unlabeled data
• merge all data and their labels and retrain the model

Question 15

Weakly-supervised learning

Answer 15

• labels in the trained model are inaccurate
• class activation maps is to calculate regional features by projecting back the weights of the output layer to convolutional feature maps

Question 16

Self-supervised learning

Answer 16

pretext task is formulated by correlating unlabeled data and its semantics