Multimodal LLMs

Question 1

What is the main limitation of context-free word embeddings like Word2Vec and GloVe?

Answer 1

They assign the same embedding to a word regardless of its context

Question 2

In transformer models, the purpose of positional encoding is to:

Answer 2

Encode the sequence order of tokens

Question 3

Which of the following is true about self-attention?

Answer 3

It compares elements within the same sequence

Question 4

Which task is BERT primarily designed for?

Answer 4

Masked Language Modeling

Question 5

What is the role of the query in a cross-attention mechanism?

Answer 5

To query information from keys and values from another modality

Question 6

In CLIP, the Image Encoder is most commonly based on which architecture?

Answer 6

ResNet or Vision Transformer (ViT)

Question 7

Which architecture uses masked self-attention to prevent tokens from seeing future tokens during training?

Answer 7

Decoder-only Transformers

Question 8

In multimodal retrieval, cosine similarity is used to:

Answer 8

Compute distances between image and text embeddings

Question 9

Which type of tokenization is typically used in modern LLMs to handle rare words efficiently?

Answer 9

Subword tokenization

Question 10

In Vision Transformers (ViT), the input image is first:

Answer 10

Split into patches and embedded

Question 11

Which of the following are advantages of using word embeddings?
A) Fixed vector length
B) Independent of vocabulary size
C) Direct interpretation of sentence grammar
D) Quantifiable semantic similarity

Answer 11

Fixed vector length; Quantifiable semantic similarity

Question 12

Which layers or blocks are typically involved in the Transformer encoder?
A) Multi-head self-attention
B) Feed-forward networks
C) Positional encoding
D) Convolutional layers

Answer 12

Multi-head self-attention; Feed-forward networks; Positional encoding

Question 13

In the Attention mechanism, which components must have the same dimensionality?
A) Queries
B) Keys
C) Values
D) Embeddings

Answer 13

Queries; Keys

Question 14

Which tasks are mentioned as applications for multimodal transformers?
A) Image Captioning
B) Vision Question Answering (VQA)
C) Text Summarization
D) Speech Recognition

Answer 14

Image Captioning; Vision Question Answering (VQA); Speech Recognition

Question 15

Which of the following are properties required for positional encoding?
A) Unique for each time step
B) Dependent on input length
C) Deterministic
D) Parallelizable

Answer 15

Unique for each time step; Deterministic; Parallelizable

Question 16

Which of the following statements about CLIP are true?
A) It uses contrastive learning to align images and text
B) It uses a shared encoder for text and image
C) It computes cosine similarity between image and text embeddings
D) It requires fine-tuning for every new task

Answer 16

It uses contrastive learning to align images and text; It computes cosine similarity between image and text embeddings

Question 17

Which techniques are used to generate embeddings?
A) CNNs
B) MLPs
C) Tokenizers
D) Autoencoders

Answer 17

CNNs; MLPs; Autoencoders

Question 18

Cross-attention in the decoder is used in which tasks?
A) Speech Recognition
B) Image Captioning
C) Multimodal Retrieval
D) Text Classification

Answer 18

Speech Recognition; Image Captioning

Question 19

In a Vision Transformer (ViT), what operations are applied to image patches?
A) Flattening
B) Linear projection to embedding
C) Sinusoidal positional encoding
D) Token masking

Answer 19

Flattening; Linear projection to embedding; Sinusoidal positional encoding

Question 20

Which are characteristics of Encoder-Decoder architectures?
A) Encoder learns contextual embeddings
B) Decoder predicts next token
C) Only works for vision tasks
D) Can use cross-attention

Answer 20

Encoder learns contextual embeddings; Decoder predicts next token; Can use cross-attention