Four questions on training, classification, overfitting, and the algorithm at the heart of every neural network.
Question 1 of 4
Question 1 of 4
During training, what actually changes inside a machine learning model?
AThe number of layers in the network
BThe weights and biases — the model's numerical parameters
CThe training data is reorganised into a better order
DThe activation functions are swapped for more suitable ones
Explanation
The architecture (number of layers, type of activation functions) is fixed before training begins. What changes are the parameters: the weights and biases attached to every connection in the network. Each training step computes a prediction, measures the error (the loss), and nudges the weights slightly in a direction that reduces that error. After millions or billions of such nudges, the weights encode a rich statistical understanding of the training data.
Question 2 of 4
Predicting whether an email is spam or not spam is an example of which type of machine learning task?
ARegression
BClustering
CClassification
DReinforcement learning
Explanation
Classification assigns data points to one of a fixed set of categories. Here the categories are "spam" and "not spam" — making this binary classification. Regression predicts a continuous number (like tomorrow's temperature or a house price). Clustering discovers natural groupings without predefined labels. Reinforcement learning trains an agent to take actions to maximise reward over time. Spam detection is one of the earliest commercial applications of machine learning and remains a classic example of the task type.
Question 3 of 4
A model scores 99% accuracy on training data but only 61% on new, unseen data. What is this problem called?
AUnderfitting
BFeature engineering
COverfitting
DRegularisation
Explanation
Overfitting happens when a model learns the training data so thoroughly that it has effectively memorised it, noise and all. It performs brilliantly on examples it has seen but fails on new data because it is fitting the quirks of the training set rather than the underlying pattern. Underfitting is the opposite: the model is too simple to capture the pattern even in training data. Regularisation is one of the techniques used to combat overfitting by penalising overly complex parameter combinations.
Question 4 of 4
Gradient descent helps a model learn by doing what, repeatedly?
ARandomly shuffling the training data until a good solution appears
BAdjusting each parameter in the direction that reduces the loss most
CComparing the model to a baseline and keeping whichever performs better
DIncreasing the learning rate whenever accuracy stops improving
Explanation
Gradient descent is the core optimisation algorithm behind almost every neural network. The gradient indicates which direction to nudge each weight in order to reduce the loss function. "Descent" captures the idea of stepping downhill in the loss landscape, making the error smaller with each pass through the training data. The size of each step is the learning rate: too large and the model overshoots the minimum; too small and training becomes impractically slow.
🎉
Phase 3 Complete
4
out of 4
You understand how models actually learn — the process that drives every modern AI system.