Week 6 · Assessment Quiz

Tabular Deep Learning & FastAI Tabular

25 multiple-choice on tabular models, entity embeddings, preprocessing and the Titanic dataset, plus 5 short-answer questions.

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

Tabular data differs from image or text data primarily because:

AIt always has fewer samples than image datasets BIt consists of structured rows and columns where each column has a specific semantic meaning and type CIt can only be used for classification, not regression DIt does not require any preprocessing before being fed to a neural network

In FastAI tabular learning, categorical variables are handled using:

AOne-hot encoding applied automatically during preprocessing BStandardisation to zero mean and unit variance CEntity embeddings — each category is mapped to a dense learned vector DBucketing into equal-width numerical bins

An entity embedding for a categorical variable:

AReplaces the category with its frequency count in the training data BMaps each unique category value to a dense learned vector that captures similarity relationships between categories CConverts the category to a one-hot binary vector of fixed width DDrops the variable if it has more than 50 unique values

TabularDataLoaders.from_df(df, ...) requires you to specify:

AThe names of categorical and continuous columns, the target variable, and the preprocessing steps (procs) BOnly the path to a CSV file — FastAI infers everything else automatically CThe neural network architecture to use for training DThe GPU memory budget for the training run

Which of the following is a continuous variable?

APassenger class (1st, 2nd, 3rd) on the Titanic BGender (Male/Female) of a passenger CAge in years of a passenger DEmbarkation port (Southampton, Cherbourg, Queenstown)

FastAI automatically applies normalisation to continuous variables because:

AContinuous variables must sum to 1.0 for the loss function to work correctly BVariables on very different scales (e.g. age vs fare) can cause unstable gradients; normalisation brings them to a common scale CNeural networks cannot process negative numbers in continuous columns DNormalisation makes the data categorical so embeddings can be applied

The FillMissing processor in FastAI:

AImputes missing values with the column median (or mode for categoricals) and adds a boolean indicator column for each imputed column BDrops all rows that contain any missing value from the dataset CFills missing values with zeros for continuous and "unknown" for categorical columns DRaises an error if any missing values are found, requiring the user to handle them manually

Categorify in FastAI:

AConverts continuous variables into equal-frequency bins and treats them as categories BApplies PCA to reduce the dimensionality of the categorical columns CConverts string categorical columns to integer codes and builds the vocabulary for each column DOne-hot encodes all categorical variables into binary dummy columns

Dropout during training works by:

ASetting all weights below a threshold to zero permanently BRandomly zeroing a fraction of activations in a layer during each forward pass, forcing the network to learn redundant representations CReducing the learning rate by a factor equal to the dropout probability DAdding Gaussian noise to the input data at the start of each epoch

Q10

layers=[200, 100] in tabular_learner creates a network with:

ATwo hidden layers: 200 neurons then 100 neurons, followed by an output layer BA single hidden layer with 200 + 100 = 300 neurons CTwo output heads: one with 200 classes and one with 100 classes D200 training epochs with a batch size of 100

Q11

The most appropriate evaluation metric for binary classification on the Titanic dataset is:

AMean Squared Error (MSE) BRoot Mean Squared Error (RMSE) CAccuracy (or F1-score if the dataset is imbalanced) DPerplexity

Q12

In FastAI tabular, cat_names and cont_names specify:

AThe output class names and the number of continuous output values BThe lists of categorical and continuous input column names respectively CThe names of the training and validation splits DThe target variable name and the list of feature columns

Q13

Batch normalisation in a neural network:

ANormalises the activations within each layer across the batch, reducing internal covariate shift and allowing higher learning rates BApplies L2 regularisation to all weight matrices in every layer CNormalises the input data once before training begins, using the training set statistics DScales the gradient to have unit norm before each weight update

Q14

A high feature importance score for a column in a tree-based model indicates:

AThe column has the highest mean value in the training data BThe column has the lowest variance and is therefore most stable CSplits on that column lead to the largest reductions in impurity across the trees, meaning it is highly predictive DThe column was used as the primary sort key during data loading

Q15

The key advantage of entity embeddings over one-hot encoding for a column with 500 unique categories is:

AOne-hot encoding requires 500 columns (sparse, high-dimensional), while an embedding of size 10 is dense and captures similarity in a compact space BEntity embeddings always achieve higher accuracy than one-hot encoding regardless of the dataset COne-hot encoding cannot be used with neural networks at all DEntity embeddings reduce the number of unique categories to at most 50

Q16

valid_idx in TabularDataLoaders specifies:

AThe list of row indices that form the validation set BThe index of the target variable column in the dataframe CThe number of validation steps per epoch DThe indices of columns to exclude from training

Q17

Class imbalance in the Titanic dataset (more non-survivors than survivors) can cause:

AThe model to train faster because there are fewer positive examples to process BThe normalisation to fail because the class means are different CA model that appears accurate by predicting the majority class for every example, while actually failing on the minority class DEmbeddings for the minority class to be initialised incorrectly

Q18

The procs argument in TabularDataLoaders specifies:

AThe list of target variable names for multi-output prediction BA list of preprocessing steps (e.g. Categorify, FillMissing, Normalize) applied to the dataset CThe number of parallel processes used for data loading DThe list of metrics to track during training

Q19

learn.predict(row) for a tabular model takes a:

AA pandas Series or single-row DataFrame representing one example, and returns (predicted class, index, probabilities) BA full DataFrame and returns predictions for all rows CA numpy array of raw feature values only, with no column names DA batch of examples as a PyTorch tensor of shape (batch, features)

Q20

Which of the following is the correct FastAI call to train a tabular learner for 5 epochs?

Alearn.tabular_fit(5) Blearn.train(n_epochs=5) Clearn.fit_one_cycle(5) Dlearn.finetune(5, pretrained=False)

Q21

In the Titanic dataset, Pclass (passenger class: 1, 2, 3) is best treated as:

AA continuous variable because it is stored as an integer BA categorical variable because the values represent distinct groups, not a continuous scale CA binary variable — collapse 2nd and 3rd class into a single "not-first" category DAn ignored variable because class is not predictive of survival

Q22

K-fold cross-validation is used to:

AObtain a more robust estimate of model performance by training and evaluating on k different splits of the data BTrain k separate models and combine their outputs into an ensemble CAutomatically tune hyperparameters by training k models with different learning rates DSplit the data into k equal classes for multi-class classification

Q23

What does tabular_learner create?

AA decision tree trained on tabular data using the CART algorithm BA linear regression model with L1 regularisation CA neural network with embedding layers for categorical variables and fully connected layers for continuous variables DA random forest with automatically tuned hyperparameters

Q24

Adding more hidden layers to a tabular neural network:

AAlways improves validation accuracy by capturing more complex patterns BIncreases model capacity — it can learn more complex relationships — but also increases the risk of overfitting CReduces training time because each layer processes fewer features DHas no effect on models trained on tabular data

Q25

Using Pclass and Sex as features to predict Titanic survival raises which ethical concern?

AThe model may perpetuate historical biases (e.g. class privilege, gender-based evacuation rules) that should not be reproduced in future decisions BThe model cannot achieve high accuracy because these features are not correlated with survival CUsing categorical features violates the assumptions of neural networks DSex cannot be stored as a categorical variable in a FastAI DataLoader

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

Q26

Explain the difference between categorical and continuous variables in tabular data. How does FastAI handle each type differently in its preprocessing pipeline?written

Your answer

0 / 700

Q27

What is entity embedding for categorical variables? Why is it more powerful than one-hot encoding for a column with many unique values (e.g. zip code with 10,000 unique values)?written

Your answer

0 / 700

Q28

Describe the three main preprocessing steps in FastAI's tabular module: FillMissing, Categorify, and Normalize. What does each do and why is each needed?written

Your answer

0 / 700

Q29

What is dropout and how does it act as a regulariser? What would you expect to happen to training and validation accuracy if you set dropout to 0.9 (very high)?written

Your answer

0 / 700

Q30

Describe how you would build a FastAI tabular model to predict Titanic survival. Include data loading, feature selection, DataLoader creation, model training, and evaluation steps.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.

Tabular Deep Learning & FastAI Tabular

Multiple Choice (25 marks)

Short Answer (5 marks — marked by lecturer)