week 2

Question 1

what is u net

Answer 1

a deep learning model for biomedical image segmentation that labels every pixel

Question 2

what does u net specialize in

Answer 2

Precise pixel-wise classification in medical images.

Question 3

What are the two paths in U-Net’s architecture?

Answer 3

Contracting path (encoder) and expanding path (decoder).

Question 4

What does the contracting path in U-Net do?

Answer 4

The “shrinking” side of U-Net that finds patterns (edges, shapes) by making the image smaller and more focused.

Question 5

What does the expanding path in U-Net do?

Answer 5

zooming in

The “growing” side that brings the image back to normal size and labels each pixel using the info it learned.

Question 6

What is the purpose of skip connections in U-Net?

Answer 6

These are shortcuts that help the decoder remember small details from the original image. Keeps things sharp.

Question 7

pixel wise classification

Answer 7

Instead of saying “this is a dog,” U-Net says “this pixel is part of a dog, this pixel is background,” etc.

Question 8

Why is U-Net good for?

Answer 8

small data sets

Question 9

data augmentation

Answer 9

Making more training images by flipping, rotating, or stretching existing ones — helpful when we don’t have a lot of real data.

Question 10

max pooling

Answer 10

A way of making the image smaller by keeping only the most important parts — like zooming out to see big patterns.

Question 11

Transposed Convolutions (Upsampling_

Answer 11

Makes the image bigger again after shrinking — used in the decoder to go back to original size.

Question 12

Cross-Entropy Loss

Answer 12

A way to measure how wrong the AI’s pixel guesses are — helps it learn and improve.

Question 13

SGD (Stochastic Gradient Descent)

Answer 13

The learning method U-Net uses — it learns step by step by checking mistakes and adjusting.

Question 14

Overlap-Tile Strategy

Answer 14

U-Net cuts up big images into tiles and looks at each one — then stitches them back together. Great for large medical scans.

Question 15

IoU (Intersection over Union)

Answer 15

A score that shows how well U-Net’s predictions match the real image. Higher = better.

Question 16

What U-Net Is Best At

Answer 16

It’s really good at medical images, especially when you don’t have a lot of labeled data.

Question 17

CLIP

Answer 17

An AI that connects images and text — it can guess what’s in a picture just from a sentence like “a photo of a cat.”

Question 18

contrastive learning

Answer 18

A way of teaching AI by showing it the right image–text pairs and telling it which ones don’t match.

Question 19

zero shot learning

Answer 19

CLIP can understand new tasks without extra training — just by reading your description.

Question 20

Natural Language Supervision

Answer 20

CLIP learns from real language (captions from the internet), not from labeled datasets like “cat = 1”

Question 21

Image Encoder

Answer 21

A part of CLIP that looks at pictures and turns them into numbers (embeddings) for the model to understand.

Question 22

Text Encoder

Answer 22

A part of CLIP that reads text and also turns it into numbers so it can be matched with the picture

Question 23

Shared Embedding Space

Answer 23

Both image and text are turned into numbers in the same space, so CLIP can match them up easily.

Question 24

Pretraining

Answer 24

CLIP is trained on 400 million image + text pairs from the internet to learn how pictures and language go together.

Question 25

Vision Transformer (ViT)

Answer 25

A powerful type of model that CLIP can use to look at pictures — better than older models for many tasks.

Question 26

Zero-Shot Classification

Answer 26

Example: You show CLIP a picture and say “Is this a dog, cat, or chair?” — it figures it out without being trained on that list.

Question 27

Why CLIP Is Useful

Answer 27

It works well on lots of tasks like reading signs, identifying objects, and even labeling images — without retraining.

Question 28

Where CLIP Struggles

Answer 28

Doesn’t do great on very specific tasks like X-rays or medical images — not trained deeply on those.

Question 29

Why CLIP Matters

Answer 29

It shows that AI can learn about the world using just language + images, like we do.

Question 30

What is the primary purpose of the U-Net architecture? A) Object detection in natural images B) Pixel-wise classification for biomedical image segmentation C) Text-to-image generation D) Speech recognition in noisy environments

Answer 30

B

Question 31

How does U-Net handle the challenge of limited labeled biomedical data? A) Using transfer learning from ImageNet B) Incorporating a generative adversarial network (GAN) C) Applying extensive data augmentation techniques D) Utilizing pre-trained transformers

Answer 31

C

Question 32

Which of the following is a key innovation introduced by U-Net? A) Fully connected layers in segmentation B) Attention mechanism for pixel grouping C) Skip connections between encoder and decoder D) Fixed input size with no padding

Answer 32

C

Question 33

What is the role of the contracting path (encoder) in U-Net? A) To classify each pixel into a specific object class B) To synthesize images from noise C) To extract and downsample features from the input image D) To compute the loss function

Answer 33

C

Question 34

How does the expanding path (decoder) work in U-Net? A) Reduces feature dimensions through convolution B) Uses transposed convolutions to upsample and combines them with encoder features C) Flattens the feature maps for classification D) Ignores previous feature maps to avoid overfitting

Answer 34

a

Question 35

Why are skip connections important in U-Net? A) They increase training time B) They improve numerical stability C) They help preserve spatial information lost during downsampling D) They reduce memory usage

Answer 35

C

Question 36

What kind of loss function is used in U-Net for pixel classification? A) Mean Squared Error B) Hinge Loss C) Pixel-wise softmax followed by cross-entropy loss D) Kullback–Leibler Divergence

Answer 36

c

Question 37

What strategy does U-Net use to segment large images efficiently?

Answer 37

c

Question 38

In the experiments, what metric did U-Net achieve the lowest value for in the EM segmentation challenge?

Answer 38

b

Question 39

Which statement best summarizes the contribution of U-Net? A) It provides a new loss function for generative models. B) It introduces reinforcement learning for segmentation tasks. C) It enables accurate and fast biomedical image segmentation using a novel architecture and augmentation. D) It is a supervised language model for image captioning.

Answer 39

c

Question 40

What is the core idea behind CLIP’s training strategy? A) Classify pixels using convolutional layers B) Predict object bounding boxes from text C) Learn visual representations by matching images with their correct text captions D) Generate images from natural language prompts

Answer 40

c

Question 41

What type of architecture does CLIP use to process text data? A) LSTM-based decoder B) Convolutional network C) Transformer-based text encoder D) Recurrent neural network (RNN)

Answer 41

c

Question 42

How does CLIP perform zero-shot classification? A) By fine-tuning on the new dataset B) By predicting pixel-level labels for each class C) By comparing image embeddings with text embeddings of class descriptions D) By retrieving nearest neighbors from a training set

Answer 42

c

Question 43

What is the main benefit of CLIP's shared embedding space? A) Enables model compression for deployment B) Allows direct comparison of images and texts for similarity C) Supports training without gradient descent D) Reduces the number of parameters in the model

Answer 43

b

Question 44

Which of the following is a limitation of CLIP mentioned in the paper? A) Inability to learn from text B) Poor performance on OCR and video classification tasks C) Inferior performance on specialized domains like medical imaging D) Requires millions of human-annotated labels

Answer 44

c