Back to Subject

Unit 4: Introduction to deep neural networks; Modern NLP - Practice Quiz

INT428 — Artificial Intelligence Essentials 60 Questions
0 Correct 0 Wrong 60 Left
0/60

1 What is a neural network primarily inspired by?

Introduction to Neural Networks Easy
A. The design of digital calculators
B. The layout of computer memory chips
C. The structure and function of the human brain
D. The organization of relational databases

2 What is the basic computational unit of a neural network called?

Introduction to Neural Networks Easy
A. Register
B. Pixel
C. Neuron (node)
D. Token

3 Who introduced the perceptron model?

Perceptron Easy
A. John McCarthy
B. Geoffrey Hinton
C. Frank Rosenblatt
D. Alan Turing

4 A single-layer perceptron can only classify data that is:

Perceptron Easy
A. Linearly separable
B. Exponentially scaled
C. Randomly distributed
D. Circularly clustered

5 What does MLP stand for in neural networks?

MLP Easy
A. Multi-Layer Perceptron
B. Managed Layer Protocol
C. Maximum Learning Process
D. Multiple Linear Predictor

6 Which layer in an MLP is located between the input and output layers?

MLP Easy
A. Hidden layer
B. Recurrent layer
C. Pooling layer
D. Convolution layer

7 CNNs are most commonly used for which type of data?

CNN Easy
A. Tabular spreadsheets
B. Images
C. SQL queries
D. Audio timestamps only

8 What does the 'C' in CNN stand for?

CNN Easy
A. Convolutional
B. Computational
C. Connected
D. Cognitive

9 What type of data are Recurrent Neural Networks (RNNs) especially suited for?

RNN Easy
A. Unrelated data points
B. Sequential data
C. Fixed-size vectors only
D. Static images

10 What feature allows an RNN to remember information from previous steps in a sequence?

RNN Easy
A. A hidden state that is passed forward through time steps and updated at each step to retain past context
B. Convolution filters
C. Max pooling
D. One-hot encoding

11 Which mechanism is the core innovation of the Transformer architecture?

Transformer Architecture and Applications Easy
A. Recurrence
B. Convolution
C. Pooling
D. Self-attention

12 What is the title of the 2017 paper that introduced the Transformer?

Transformer Architecture and Applications Easy
A. Long Short-Term Memory
B. Attention Is All You Need
C. ImageNet Classification
D. Deep Residual Learning

13 What does NLP stand for?

Introduction to NLP Easy
A. Numeric Language Parsing
B. Neural Learning Protocol
C. Natural Language Processing
D. Nested Loop Programming

14 Which NLP phase deals with the grammatical structure of sentences?

NLP phases Easy
A. Phonetic analysis
B. Discourse analysis
C. Syntactic analysis
D. Pragmatic analysis

15 What is tokenization in NLP?

Tokenization Easy
A. Encrypting text into cipher form
B. Translating text into another language
C. Converting text into audio
D. Breaking text into smaller units such as words or subwords

16 What do word embeddings represent?

Embeddings Easy
A. Words as audio waveforms
B. Words as raw ASCII codes
C. Words as dense numerical vectors capturing meaning
D. Words as fixed image pixels

17 What does the attention mechanism help a model do?

Attention Easy
A. Convert text into images
B. Increase the file size of the model
C. Reduce the total number of layers
D. Focus on the most relevant parts of the input

18 What does the 'G' in GPT stand for?

Language models (BERT, GPT) Easy
A. Grouped
B. Guided
C. Graphical
D. Generative

19 BERT is designed to read text in which manner?

Language models (BERT, GPT) Easy
A. One character at a time
B. Bidirectionally, considering both left and right context of a word at the same time to understand meaning
C. Only right to left
D. Only left to right

20 Which NLP task involves determining whether a piece of text is positive, negative, or neutral?

NLP use cases (sentiment analysis, translation, summarization) Easy
A. Machine translation
B. Named entity recognition
C. Text summarization
D. Sentiment analysis

21 A single perceptron uses a step activation and weights , , bias . For input , what is the output if the threshold is at ?

Perceptron Medium
A. 1
B. 0.5
C. 0
D. -0.2

22 Why can a single-layer perceptron not solve the XOR problem?

Perceptron Medium
A. XOR is not linearly separable
B. XOR uses continuous outputs only
C. XOR requires more than two inputs
D. XOR needs a sigmoid activation

23 During backpropagation, what is primarily computed to update the network weights?

Introduction to Neural Networks Medium
A. Second derivative of the activation
B. Average of all layer outputs
C. Variance of the input features
D. Gradient of the loss with respect to each weight

24 Which property of non-linear activation functions makes a multi-layer perceptron more powerful than a single linear layer?

MLP Medium
A. They let the network approximate non-linear functions
B. They reduce the number of parameters needed
C. They guarantee faster convergence always
D. They remove the need for a bias term

25 A fully connected layer maps 128 inputs to 64 outputs. Ignoring bias, how many weight parameters does it contain?

MLP Medium
A. 192
B. 2048
C. 4096
D. 8192

26 Applying a convolution filter with stride 1 and no padding to a input produces an output of what spatial size?

CNN Medium
A.
B.
C.
D.

27 What is the main purpose of a pooling layer in a CNN?

CNN Medium
A. Normalize the input pixel values
B. Reduce spatial dimensions and add invariance
C. Introduce non-linearity into the model
D. Increase the number of channels

28 Why is parameter sharing in convolutional layers advantageous compared to fully connected layers for images?

CNN Medium
A. It removes the need for activation functions
B. It allows variable-length text inputs
C. It guarantees the network cannot overfit
D. The same filter detects a feature anywhere in the image with fewer parameters

29 What problem do LSTM and GRU units specifically address in standard RNNs?

RNN Medium
A. The absence of a loss function
B. The lack of any recurrent connection
C. The vanishing gradient over long sequences
D. The inability to process numeric data

30 In an RNN, what makes it suitable for sequential data compared to a plain MLP?

RNN Medium
A. It maintains a hidden state passed across time steps
B. It uses convolution over the sequence
C. It processes all tokens fully in parallel
D. It requires no weight training

31 Why do Transformers require positional encodings to be added to input embeddings?

Transformer Architecture and Applications Medium
A. Embeddings cannot store word meaning
B. Positional encodings replace the need for weights
C. Attention cannot handle more than 512 tokens
D. Self-attention itself is order-agnostic

32 What is a key computational advantage of Transformers over RNNs when training on long sequences?

Transformer Architecture and Applications Medium
A. They need no attention mechanism
B. Tokens can be processed in parallel
C. They avoid using any matrix multiplication
D. They use far fewer parameters

33 In scaled dot-product attention, the score is divided by mainly to:

Attention Medium
A. Prevent large dot products from saturating the softmax
B. Reduce the number of attention heads
C. Guarantee the weights sum to zero
D. Convert scores into embeddings

34 What is the primary benefit of using multi-head attention instead of a single attention head?

Attention Medium
A. It reduces the sequence length automatically
B. It eliminates the positional encoding requirement
C. It lets the model attend to different representation subspaces jointly
D. It removes the need for feed-forward layers

35 Why do modern language models often use subword tokenization (e.g. Byte-Pair Encoding) instead of whole-word tokenization?

Tokenization Medium
A. It only works for numeric inputs
B. It removes the need for embeddings
C. It handles rare and unknown words while limiting vocabulary size
D. It always produces exactly one token per sentence

36 In word embeddings, why can the relationship hold?

Embeddings Medium
A. Semantic relationships are captured as directions in vector space
B. Embeddings are randomly assigned and fixed
C. Words are stored as one-hot vectors
D. Each word maps to a single scalar value

37 Which NLP phase is primarily concerned with the grammatical structure and arrangement of words in a sentence?

NLP phases Medium
A. Discourse integration
B. Syntactic analysis
C. Pragmatic analysis
D. Lexical analysis

38 Consider the sentences "I saw the man with a telescope." Which NLP challenge does this best illustrate?

Introduction to NLP Medium
A. Ambiguity in interpretation
B. Missing stop words
C. Character encoding errors
D. Tokenization failure

39 What is a key architectural difference between BERT and GPT?

Language models (BERT, GPT) Medium
A. BERT has no attention mechanism
B. BERT is bidirectional while GPT is left-to-right
C. BERT is trained without any data
D. GPT cannot generate text

40 BERT's pre-training uses a Masked Language Model objective. What does this involve?

Language models (BERT, GPT) Medium
A. Predicting randomly hidden tokens from surrounding context
B. Removing all punctuation from text
C. Translating sentences into another language
D. Predicting the next word one at a time

41 A single-layer perceptron with a step activation function is trained on the XOR dataset. After many epochs, the training loss fails to reach zero. What is the fundamental reason?

Perceptron Hard
A. The step function's zero gradient prevents any weight update
B. XOR is not linearly separable, so no single hyperplane can classify all four points
C. The dataset is too small for the perceptron to generalize
D. The learning rate is too high, causing oscillation around the minimum

42 Consider an MLP where every hidden neuron uses a linear (identity) activation function across all layers. What is the expressive power of this network?

MLP Hard
A. It collapses to an equivalent single linear transformation regardless of depth
B. It can approximate any continuous function given enough neurons
C. It can learn XOR because of the multiple layers
D. It behaves like a deep non-linear classifier due to weight stacking

43 During backpropagation in a deep sigmoid network, gradients in early layers become extremely small. Which statement best explains this vanishing gradient phenomenon?

Introduction to Neural Networks Hard
A. The chain rule adds gradients, causing them to average toward zero
B. Large learning rates saturate the sigmoid at its extremes
C. Sigmoid outputs are always positive, forcing weights to grow unbounded
D. Repeated multiplication of derivatives bounded by shrinks gradients exponentially with depth

44 A CNN layer has an input of size , uses a kernel with stride and no padding. What is the spatial dimension of the output feature map?

CNN Hard
A.
B.
C.
D.

45 Why does parameter sharing in convolutional layers provide an advantage over fully connected layers for image data?

CNN Hard
A. It removes the vanishing gradient problem in deep networks
B. It eliminates the need for pooling layers entirely
C. It guarantees rotation invariance of learned features
D. It drastically reduces parameters and enforces translation-equivariant feature detection

46 An LSTM is preferred over a vanilla RNN for long sequences primarily because:

RNN Hard
A. It uses fewer parameters and trains faster on all tasks
B. It replaces the tanh activation with ReLU to avoid saturation
C. Its gating and additive cell state allow gradients to flow without vanishing across many steps
D. It processes the entire sequence in parallel like a Transformer

47 In scaled dot-product attention, the dot products of queries and keys are divided by . What problem does this scaling primarily address?

Transformer Architecture and Applications Hard
A. It ensures the attention weights sum to one across the sequence
B. It normalizes the output to a valid probability distribution
C. It reduces the computational cost of the matrix multiplication
D. For large , large dot products push softmax into regions with vanishingly small gradients

48 Why do Transformers require explicit positional encodings while RNNs do not?

Transformer Architecture and Applications Hard
A. RNNs cannot process sequences without positional signals either
B. Transformers use softmax which discards magnitude information
C. Self-attention is permutation-invariant and has no inherent notion of token order
D. Positional encodings replace the need for an embedding layer

49 Multi-head attention with heads splits the model dimension so each head operates on . What is the main benefit of using multiple heads instead of one large head?

Attention Hard
A. It makes attention weights deterministic and non-learnable
B. It reduces total computation to of single-head attention
C. It removes the need for the feed-forward sublayer
D. Different heads can jointly attend to information from different representation subspaces

50 BERT and GPT differ fundamentally in their pretraining objectives and attention. Which statement correctly contrasts them?

Language models (BERT, GPT) Hard
A. GPT uses bidirectional attention; BERT uses causal masking
B. BERT uses bidirectional attention with masked LM; GPT uses causal (left-to-right) attention with next-token prediction
C. BERT is autoregressive; GPT uses masked language modeling
D. Both use identical bidirectional attention but differ only in tokenizer

51 Subword tokenization schemes like Byte-Pair Encoding (BPE) are preferred over word-level tokenization mainly because they:

Tokenization Hard
A. Eliminate the need for an embedding matrix
B. Guarantee one token per morpheme in every language
C. Handle rare and out-of-vocabulary words by decomposing them into known subword units
D. Always produce shorter sequences than word tokenization

52 A key limitation of static word embeddings like Word2Vec compared to contextual embeddings from BERT is that static embeddings:

Embeddings Hard
A. Require a Transformer encoder to be computed
B. Assign a single fixed vector per word, ignoring polysemy across contexts
C. Cannot be trained on large corpora efficiently
D. Produce embeddings that violate the triangle inequality

53 In Word2Vec, the famous analogy works because the embedding space captures:

Embeddings Hard
A. Character-level spelling similarities between words
B. Exact one-hot encodings of each word
C. Frequency counts of word co-occurrence only
D. Consistent linear directions encoding semantic relationships

54 Consider the pipeline: raw text tokens POS tags parse tree meaning representation. Which classical NLP phase corresponds to constructing the parse tree?

NLP phases Hard
A. Pragmatic analysis
B. Discourse integration
C. Syntactic analysis (parsing)
D. Lexical analysis

55 The sentence "The trophy did not fit in the suitcase because it was too big" requires resolving what "it" refers to. This challenge is primarily handled by which NLP capability?

NLP phases Hard
A. Tokenization of pronouns
B. Coreference resolution using semantic and pragmatic reasoning
C. Morphological analysis
D. Part-of-speech tagging

56 In abstractive summarization compared to extractive summarization, the model:

NLP use cases (sentiment analysis, translation, summarization) Hard
A. Requires no language generation capability
B. Only removes stop words from the original text
C. Selects and concatenates the most important source sentences verbatim
D. Generates novel phrasing that may not appear verbatim in the source

57 A sentiment classifier labels "This movie is not bad at all" as negative. What linguistic phenomenon most likely caused this error?

NLP use cases (sentiment analysis, translation, summarization) Hard
A. The softmax temperature was miscalibrated
B. The embedding dimension was too large
C. The model failed to handle negation and its scope over the sentiment word
D. Tokenization split the contractions incorrectly

58 A retrieval-augmented (RAG) chatbot is chosen over a purely generative LLM chatbot for an enterprise knowledge base primarily to:

Building chatbots and digital assistants Hard
A. Eliminate the need for any language model
B. Reduce the token embedding dimension for faster inference
C. Guarantee grammatically perfect output in all languages
D. Ground responses in retrieved documents, reducing hallucination and enabling up-to-date facts

59 The self-attention mechanism has a computational and memory complexity that scales as which function of sequence length ?

Transformer Architecture and Applications Hard
A. linear in sequence length
B. log-linear
C. quadratic in sequence length
D. constant regardless of length

60 A bidirectional RNN improves over a unidirectional RNN for sequence tagging tasks because it:

RNN Hard
A. Encodes each token using both past and future context in the sequence
B. Can generate text autoregressively during inference
C. Trains with half the parameters of a unidirectional RNN
D. Removes recurrence, enabling full parallel computation
Previous MCQ
Unit 3
Next MCQ
Unit 5