Deep Learning — Module Assessment

Question 1

What best describes the difference between deep learning and much of classical machine learning?

Accepted Answer

Deep learning learns useful features automatically, while classical methods often rely on hand-engineered features.

Answer

Deep learning avoids using any data during training.

Answer

Deep learning always uses fewer layers than classical models.

Answer

Deep learning only works on small, tidy spreadsheets.

Question 2

Which pair of developments most directly enabled the recent rise of deep learning?

Accepted Answer

Large labeled datasets and powerful GPUs.

Answer

Smaller screens and faster internet browsers.

Answer

New programming languages and cheaper keyboards.

Answer

Fewer training examples and slower processors.

Question 3

What does an artificial neuron do with its inputs?

Accepted Answer

It multiplies each input by a weight, sums the results, and passes the total through an activation step.

Answer

It stores each input permanently without changing it.

Answer

It sends every input straight to the output layer unchanged.

Answer

It randomly picks one input and ignores the rest.

Question 4

Who introduced the perceptron, an early model of an artificial neuron, and when?

Accepted Answer

Frank Rosenblatt in 1958.

Answer

Alan Turing in 1936.

Answer

Geoffrey Hinton in 2012.

Answer

Alex Krizhevsky in 2009.

Question 5

During training, what is the direct purpose of the loss?

Accepted Answer

To measure how far the prediction is from the correct answer

Answer

To store the network's final answer for later reuse

Answer

To decide how many layers the network should have

Answer

To convert the input image into numbers

Question 6

What does backpropagation actually do after the loss is calculated?

Accepted Answer

It sends the error backward through the layers so each weight learns how much it contributed to the mistake

Answer

It replaces every weight with a fresh random number

Answer

It permanently stores the training images inside the network

Answer

It increases the learning rate until the loss reaches zero

Question 7

In a convolutional neural network, how does understanding of an image typically build up across the layers?

Accepted Answer

Early layers detect simple edges, and deeper layers combine them into shapes and then whole objects

Answer

The final layer detects edges and the first layer names the whole object

Answer

Every layer independently tries to name the whole object

Answer

The network compares the photo pixel-by-pixel against stored images

Question 8

Why is AlexNet's result at the 2012 ImageNet competition considered a turning point?

Accepted Answer

A deep convolutional network cut the error rate far below the older hand-engineered methods, showing deep learning's power for vision

Answer

It was the first computer program ever to recognize any image

Answer

It proved neural networks understand images exactly as humans do

Answer

It removed the need for any training data

Question 9

What problem with recurrent neural networks did attention most directly help solve?

Accepted Answer

Information from early words tended to fade as it was passed through many sequential memory-update steps.

Answer

They could not represent words as numbers at all.

Answer

They required no training data to work.

Answer

They could only ever process images, not text.

Question 10

Which statement best describes what the transformer architecture changed about processing sequences?

Accepted Answer

It removed strict word-by-word recurrence and used attention so the whole sequence could be processed in parallel.

Answer

It made models read text strictly one letter at a time for accuracy.

Answer

It replaced neural networks with simple keyword lookup tables.

Answer

It eliminated the need to consider the relationships between words.

Question 11

What is an adversarial example in deep learning?

Accepted Answer

An input altered by a tiny, often imperceptible change that is engineered to make the model misclassify it.

Answer

A training example that a human labelled incorrectly by accident.

Answer

A rare input that no model has ever been trained on before.

Answer

A second model built specifically to compete against the first.

Question 12

Why is the 'black-box' nature of deep learning a concern in high-stakes settings like medicine?

Accepted Answer

The reasoning is buried in millions of numerical weights, so the model usually cannot give a readable reason for its decision.

Answer

The models run only on black-coloured hardware that is hard to obtain.

Answer

The models refuse to produce any output when the stakes are high.

Answer

The models can only be used at night when systems are offline.

Question 13

Which statement best describes what AlphaGo's 2016 victory over Lee Sedol demonstrated?

Accepted Answer

That deep learning plus reinforcement learning could master one complex game at a world-class level.

Answer

That machines had achieved general, human-level intelligence.

Answer

That Go is a simple game once computers try it.

Answer

That neural networks understand strategy the way humans do.

Question 14

Why does a deep-learning model that scores well on a benchmark sometimes fail in the real world?

Accepted Answer

The model may meet data unlike its training set and fail while still reporting confident scores.

Answer

Benchmarks are always faked by the developers.

Answer

Real-world use never involves any data at all.

Answer

Models forget their training the moment they are deployed.