Question 1

What do weights determine in a neural network?

A. How important each input feature is B. The number of layers in the network C. The learning rate of the optimiser D. The size of the training dataset

Question 2

What does the bias term do in a neural network layer?

A. It multiplies each input by a learned weight B. It shifts the output up or down, allowing the model to fit data more flexibly C. It converts logits into probabilities that sum to 1 D. It controls how much the weights are updated during training

Question 3

What is the purpose of activation functions in a neural network?

A. To reduce the number of parameters B. To normalise the input data C. To introduce non-linearity, enabling the network to learn complex patterns D. To split data into training and test sets

Question 4

What does the softmax function do?

A. Squashes values between -1 and 1 B. Converts raw scores (logits) into probabilities that sum to 1 C. Removes negative values from the output D. Reduces the dimensionality of the input

Question 5

The sigmoid function squashes values into what range?

A. -1 to 1 B. 0 to 1 C. 0 to infinity D. -infinity to infinity

Question 6

Which loss function is used for binary classification?

A. Mean Squared Error (MSE) B. Mean Absolute Error (MAE) C. Binary Cross-Entropy D. Sparse Categorical Cross-Entropy

Question 7

Which regression loss function is more robust to outliers?

A. Mean Squared Error (MSE) B. Mean Absolute Error (MAE) C. Binary Cross-Entropy D. Categorical Cross-Entropy

Question 8

What happens when the learning rate is too high?

A. The model learns very slowly and gets stuck in local minima B. The model overshoots optimal weights and training becomes unstable or diverges C. The model converges efficiently to a good solution D. The model ignores the bias term

Question 9

What is Adam in the context of training neural networks?

A. A type of activation function B. A loss function for regression C. The most common optimiser that automatically adjusts the learning rate per parameter D. A method for fine-tuning LLMs

Question 10

What does a lower temperature do to model predictions?

A. Makes outputs more deterministic B. Makes outputs more diverse and random C. Increases the learning rate D. Reduces the number of attention heads

Question 11

CNNs are commonly used for which type of tasks?

A. Time series prediction B. Computer vision tasks C. Machine translation D. Text generation

Question 12

What is the role of pooling layers in a CNN?

A. Extract features from images B. Reduce spatial size and computation C. Produce final predictions D. Introduce non-linearity

Question 13

What key property allows RNNs to process sequential data?

A. Convolution layers B. Pooling layers C. Memory — they share weights and biases across time steps D. Multi-head attention

Question 14

What does positional encoding do in a Transformer?

A. Converts words into dense semantic vectors B. Lets the model understand the order of tokens in a sequence C. Reduces the sequence length for faster processing D. Removes stop words from the input

Question 15

BERT is trained with which method?

A. Masked Language Modelling (MLM) — predicting masked tokens using full surrounding context B. Causal language modelling — predicting the next token left-to-right C. Reinforcement learning from human feedback (RLHF) D. Contrastive learning on image-text pairs

Question 16

What does self-attention allow a model to do?

A. Process tokens one at a time in sequence B. Process an entire sequence simultaneously and learn dependencies between all tokens C. Only attend to adjacent tokens D. Reduce the vocabulary size

Question 17

In multi-head attention, what happens to the outputs of all heads?

A. They are averaged into a single vector B. Only the best head's output is kept C. They are concatenated and passed through a linear projection D. They are summed element-wise and normalised

Question 18

What is the main problem with one-hot encoding?

A. No sense of similarity between words — vectors are sparse and high-dimensional B. It requires a GPU to compute C. It can only represent bigrams, not single words D. It changes the vector for each word depending on context

Question 19

What is the main limitation of Word2Vec / GloVe embeddings?

A. They are sparse and high-dimensional B. They ignore word order completely C. One fixed vector per word regardless of context D. They cannot capture semantic relationships

Question 20

What makes contextual embeddings (BERT, GPT) different from Word2Vec?

A. They use one-hot encoding internally B. They produce the same vector for every occurrence of a word C. Each word gets a different vector depending on its surrounding context D. They only work with single-word inputs

Question 21

What does Bag of Words (BoW) ignore?

A. Word order B. Word frequency C. Vocabulary size D. Document length

Question 22

How does TF-IDF improve on Bag of Words?

A. It captures semantic meaning between synonyms B. It weights words by how frequent they are in a document vs. how common across all documents C. It preserves word order in the document D. It produces dense low-dimensional vectors

Question 23

What is chunking in the context of retrieval systems?

A. Training a model on smaller batches of data B. Splitting large documents into smaller pieces before embedding and retrieval C. Compressing model weights to reduce memory usage D. Breaking a neural network into parallel sub-networks

Question 24

Which of the following is a parameter-efficient LLM fine-tuning method?

A. Full Fine-Tuning B. LoRA / PEFT C. Adam optimiser D. Softmax

Question 25

How many bytes per parameter does FP16 use?

A. 1 byte B. 2 bytes C. 4 bytes D. 8 bytes

Question 26

Attention is particularly important in which type of architecture?

A. Convolutional Neural Networks B. Encoder-decoder architectures such as machine translation C. Random forests D. K-nearest neighbours

Question 27

Given the sentence "I love NLP", what are the bigrams?

A. [I, love, NLP] B. [I love, love NLP] C. [I love NLP] D. [I NLP, love NLP]

Question 28

How does Sparse Categorical Cross-Entropy differ from Categorical Cross-Entropy?

A. It is used for regression instead of classification B. It only works with binary labels C. It accepts integer class labels instead of one-hot vectors D. It penalises large errors more heavily

Question 29

Why is multi-head attention more powerful than a single attention head?

A. It uses fewer parameters B. It processes tokens one at a time for accuracy C. Multiple heads capture several different relationships simultaneously D. It removes the need for positional encoding

Question 30

What does Recursive Character Level Chunking do?

A. Splits text at a fixed character length regardless of content B. Recursively splits on a hierarchy of separators (paragraphs, sentences, words) until chunks are small enough C. Uses an LLM to determine split points based on semantic context D. Uses the document's headings and sections to define boundaries

Question 31

What causes the vanishing gradient problem?

A. Gradients grow unboundedly large during backpropagation B. The learning rate is set too high C. Gradients become extremely small as they propagate back through many layers, causing early layers to stop learning D. The model has too few parameters to fit the data

Question 32

What is the primary solution for the exploding gradient problem?

A. Switching from ReLU to sigmoid activations B. Adding more layers to the network C. Gradient clipping — capping gradients at a maximum value D. Increasing the learning rate

Question 33

Which activation functions are most likely to cause the vanishing gradient problem?

A. ReLU and Leaky ReLU B. Sigmoid and Tanh C. Softmax and ReLU D. Leaky ReLU and Tanh

Question 34

Which of the following is NOT a key limitation of vanilla RNNs?

A. Vanishing and exploding gradients B. Short-term memory — losing context from early time steps C. Inability to process sequential data D. Sequential processing that cannot be parallelised

Question 35

What do LSTM and GRU architectures solve compared to vanilla RNNs?

A. They enable GPU parallelism for faster training B. They address vanishing gradients and short-term memory using gating mechanisms C. They remove the need for activation functions D. They replace recurrence with self-attention

Question 36

What is the main bottleneck of the original seq2seq encoder-decoder architecture?

A. The decoder cannot generate variable-length outputs B. The entire input is compressed into a single fixed-size context vector, losing information on long sequences C. It can only translate between two specific languages D. The encoder and decoder must have the same number of layers

Question 37

How does the attention mechanism (Bahdanau, 2015) fix the seq2seq bottleneck?

A. By removing the encoder entirely B. By doubling the size of the context vector C. By letting the decoder look back at all encoder hidden states and compute a weighted context at every step D. By processing the input in reverse order

Question 38

What does the KV cache store during autoregressive generation?

A. The model's weights and biases for each layer B. The Query vectors from all previous tokens C. The Key and Value matrices from all previous tokens so they don't need to be recomputed D. The softmax probabilities of previously generated tokens

Question 39

What is the main trade-off of using a KV cache?

A. It trades extra GPU memory for significantly faster inference B. It reduces model accuracy in exchange for speed C. It requires retraining the model with a different architecture D. It only works with encoder-only models like BERT

Question 40

What does Grouped-Query Attention (GQA) do to reduce KV cache memory?

A. It removes the Value matrices entirely B. Groups of attention heads share the same Key and Value projections C. It halves the number of Transformer layers D. It quantises the Query vectors to INT4

Question 41

What effect does using a smaller batch size have on training?

A. More stable gradients and faster convergence B. Noisier gradient updates that can act as regularisation C. Higher GPU memory usage D. It eliminates the need for dropout

Question 42

What does dropout do during training?

A. Removes entire layers from the network B. Randomly sets a fraction of neurons to zero each step, preventing over-reliance on any single neuron C. Reduces the learning rate over time D. Clips gradients to prevent exploding gradients

Question 43

In a fraud detection system, 99.5% of transactions are legitimate. If a model predicts "not fraud" for every transaction, what is its accuracy?

A. 0% B. 50% C. 99.5% — but it catches zero fraud, so accuracy is misleading with class imbalance D. It depends on the loss function

Question 44

In fraud detection, which is typically more costly to get wrong — a false positive (flagging a legitimate transaction) or a false negative (missing actual fraud)?

A. False positive — it causes the most financial damage B. False negative — missing real fraud means direct financial loss, so high recall is critical C. They are equally costly D. Neither matters if accuracy is above 95%

Question 45

What is Retrieval-Augmented Generation (RAG)?

A. A fine-tuning method that trains the model on retrieved documents B. A technique that retrieves relevant documents and feeds them as context to the LLM at inference time to ground its responses C. A type of attention mechanism that retrieves past hidden states D. A method of compressing model weights for faster inference

Question 46

When should you use RAG instead of fine-tuning?

A. When the knowledge changes frequently and needs to stay up-to-date without retraining B. When you want to change the model's writing style or tone C. When you have no external documents to retrieve from D. When the model needs fewer parameters

Question 47

What is a hallucination in the context of LLMs?

A. When the model takes too long to generate a response B. When the model generates fluent but factually incorrect or fabricated information C. When the model repeats the same token indefinitely D. When the model refuses to answer a question

Question 48

Named Entity Recognition (NER) is used in banking KYC processes to:

A. Generate new customer documents automatically B. Extract structured information like names, addresses, and ID numbers from unstructured documents C. Translate documents between languages D. Compress documents for storage

Question 49

A 13B parameter model in FP32 (4 bytes per parameter) requires 52 GB. How much memory does the same model need in INT8 (1 byte per parameter)?

A. 26 GB B. 13 GB C. 6.5 GB D. 104 GB

Question 50

How does LoRA achieve parameter-efficient fine-tuning?

A. It removes half the layers from the model before training B. It freezes the model and only trains the embedding layer C. It freezes the original weights and injects small trainable low-rank matrices alongside them D. It reduces the vocabulary size to lower memory usage