Week 2 · Assessment Quiz

ML Fundamentals & DataBlock API

25 multiple-choice on overfitting, metrics, loss functions and the FastAI DataBlock API, plus 5 short-answer questions.

📋 30 questions total ⭐ 30 marks 🕐 No time limit 🔒 Answers not revealed

Overfitting occurs when:

AThe model is too simple to capture patterns in the training data BThe model learns the training data too well and fails to generalise to unseen examples CThe learning rate is too small to update weights effectively DThe validation set is larger than the training set

The primary purpose of a validation set is to:

AAdd more data to gradient updates BProvide a final unbiased assessment after all training is complete CMonitor model performance during training to guide decisions like early stopping DAugment the training set with synthetic examples

In FastAI, the DataBlock API is used to:

ADefine the neural network architecture BSpecify how to get, split, label, and transform data in a reusable pipeline CDownload datasets from the internet automatically DTrack and log training metrics to a dashboard

Which of the following is an example of a classification task?

APredicting whether an email is spam or not spam BPredicting the price of a house given its features CEstimating tomorrow's maximum temperature in degrees DPredicting the number of customers who will visit a store

The difference between a training set and a test set is:

ATraining sets are larger; test sets are used to improve training BTraining sets use augmentation; test sets do not CThe model is trained on the training set; the test set is used only once for final evaluation DTest sets are labelled; training sets are unlabelled

Cross-entropy loss is most suitable for:

ARegression tasks predicting a continuous numeric output BClassification tasks where outputs are probability distributions over classes CRanking tasks where items must be ordered by relevance DReconstruction tasks in autoencoders

A learning rate that is too small causes:

AVery slow convergence — training takes many more epochs to reach a good solution BThe loss to diverge and increase each step CWeights to be updated in the wrong direction DNo gradient updates at all — the model stays at its initialisation

RandomSplitter(valid_pct=0.2) in FastAI:

ARandomly augments 20% of the training images BDrops 20% of the dataset to speed up training CRandomly assigns 20% of the data to the validation set and the rest to training DSplits each image into 20% overlapping patches

Accuracy as a metric is defined as:

ATotal number of correct predictions divided by the number of classes BNumber of correct predictions divided by total number of predictions C1 minus the loss value on the validation set DThe probability assigned to the true class averaged over all samples

Q10

Supervised learning differs from unsupervised learning in that:

AIt requires labelled training data where each input has a known output BIt does not require any data — the model learns from simulated environments CIt only works for image data, not text or tabular data DIt uses clustering to find natural groupings in the data

Q11

Which of the following is NOT a common data augmentation technique for images?

ARandom horizontal flipping BRandom brightness/contrast adjustment CReplacing all pixel values with their logarithm DRandom cropping and resizing

Q12

Batch size determines:

AThe number of layers in the neural network BHow many training examples are used to compute each gradient update CThe maximum number of epochs during training DHow many classes the model can predict

Q13

A model with 99% training accuracy and 55% validation accuracy is most likely:

AUnderfitting — the model is too simple for the task BWell-regularised — this gap is expected and acceptable COverfitting — the model has memorised the training data but cannot generalise DUsing too large a learning rate — loss is diverging

Q14

In the FastAI DataBlock, blocks= specifies:

AThe number of hidden layers in the model BThe types of input and output (e.g. ImageBlock and CategoryBlock) CThe batch size and number of workers for loading DThe data augmentation pipeline to apply

Q15

aug_transforms() in FastAI provides:

AA standard set of image augmentations applied per batch (flips, zoom, warp, lighting) BA way to download augmented copies of the dataset from the internet CA set of text augmentations for NLP tasks DAugmentation applied to model weights during training

Q16

A categorical variable in machine learning is:

AAny variable measured in whole numbers BA variable that takes discrete named values (e.g. colour: red, green, blue) CA variable with values in the range [0, 1] DA variable that represents a physical measurement

Q17

Which loss function is appropriate for multi-class classification?

AMean squared error (MSE) BBinary cross-entropy CCategorical cross-entropy (softmax + negative log likelihood) DHuber loss

Q18

Image normalisation in deep learning typically involves:

AScaling pixel values to [0, 255] BSubtracting the dataset mean and dividing by the standard deviation per channel CConverting all images to greyscale before training DApplying Gaussian blur to reduce high-frequency noise

Q19

In the DataBlock API, get_items is responsible for:

AReturning the list of all raw items (e.g. file paths) from which the dataset is built BApplying transformations to each individual item CSplitting items into training and validation sets DCreating the DataLoader from the processed dataset

Q20

item_tfms in FastAI applies transforms:

AOnce per batch, on a GPU, in parallel BAt model inference time only, not during training CTo each individual item before batching (e.g. resize to a fixed size) DTo the model weights rather than the data

Q21

For images organised into class-named folders, which FastAI shortcut is most convenient?

ADataBlock with custom get_items function BImageDataLoaders.from_folder() CTabularDataLoaders.from_df() DCollabDataLoaders.from_csv()

Q22

Underfitting is best identified when:

ABoth training and validation accuracy are low BTraining accuracy is much higher than validation accuracy CThe model takes too long to train DThe validation loss is exactly equal to the training loss

Q23

What does a softmax output layer produce?

ARaw scores (logits) for each class with no constraint on their sum BA probability distribution over all classes that sums to 1.0 CA binary 0 or 1 prediction for each class independently DThe class with the highest score, discarding all probabilities

Q24

Which of the following best describes a regression task?

AAssigning an image to one of 10 fixed categories BPredicting whether a transaction is fraudulent (yes/no) CPredicting a continuous numeric value such as house price DGrouping unlabelled data into natural clusters

Q25

Increasing batch size generally:

AAlways improves model accuracy BReduces memory usage per step and may reduce generalisation quality CSpeeds up training at the cost of slower convergence DHas no effect on training dynamics

Answer each question in 2–4 sentences. Precise technical language is expected. Code snippets are welcome where relevant.

Q26

Explain the difference between overfitting and underfitting. How can each be detected using training and validation metrics?written

Your answer

0 / 700

Q27

What is the DataBlock API in FastAI? Describe its key components (blocks, get_items, splitter, get_y, item_tfms, batch_tfms) and what each does.written

Your answer

0 / 700

Q28

What is cross-entropy loss? Why is it preferred over mean squared error for classification tasks?written

Your answer

0 / 700

Q29

Describe what item_tfms and batch_tfms do in the FastAI DataBlock API and at what stage of the pipeline each is applied.written

Your answer

0 / 700

Q30

What is data augmentation? Give three specific examples of image augmentation techniques and explain why each helps improve model generalisation.written

Your answer

0 / 700

Your full name

Complete all 30 questions then click Submit. Your MCQ score (25/25) will be shown. Short answers are marked separately.

ML Fundamentals & DataBlock API

Multiple Choice (25 marks)

Short Answer (5 marks — marked by lecturer)