8 of the 30 questions in this set, with the correct answer marked and every option explained.
1. What is deep learning primarily inspired by?
✓The structure and function of the human brain
Deep learning is inspired by the structure and function of the human brain, specifically the neural networks composed of interconnected neurons. Artificial neural networks mimic this structure to learn from data.
✗Statistical models
Deep learning is thoroughly statistical in its mathematics, so this is not wrong so much as beside the point. The architectural metaphor — layers of connected units that fire based on weighted inputs — comes from neuroscience.
✗Rule-based expert systems
Symbolic AI, where humans hand-code explicit if-then rules. This is the paradigm deep learning largely displaced, by learning the rules from data instead.
✗Evolutionary algorithms
Inspired by natural selection, using mutation and selection over populations. A different biological metaphor entirely, though the two are occasionally combined in neuroevolution.
2. Which of the following is a characteristic of a deep neural network?
✓It has multiple layers between input and output layers.
A deep neural network is defined by having multiple hidden layers between the input and output layers, allowing it to model complex, hierarchical representations of data.
✗It relies solely on linear transformations.
A stack of purely linear layers is mathematically equivalent to a single linear layer, however many you add. Depth only buys anything when non-linearities sit between the layers.
✗It does not use activation functions.
Same problem stated differently. Remove the activations and the whole network collapses to one linear map.
✗It has only one hidden layer.
That is a shallow network. The word 'deep' refers precisely to having many hidden layers, which is what lets the model build features on top of features.
3. Which activation function is known for mitigating the vanishing gradient problem and is defined as f(x) = x for x > 0 and f(x) = α(e^x - 1) for x ≤ 0?
✓ELU (Exponential Linear Unit)
The Exponential Linear Unit (ELU) activation function helps mitigate the vanishing gradient problem by allowing a small, non-zero gradient when the unit is inactive, and it tends to converge faster and produce more accurate results.
✗ReLU
Defined as max(0, x), so it is exactly zero for all negative inputs. That flat region is what causes the dying-ReLU problem, which ELU's exponential negative branch is designed to avoid.
✗Leaky ReLU
Very close — it also permits small negative outputs. But it uses the straight line αx below zero, whereas the formula here curves exponentially as α(e^x − 1).
✗Sigmoid Function
Saturates at both extremes, so gradients shrink toward zero for large positive or negative inputs. It causes the vanishing gradient problem rather than mitigating it.
4. What is the primary advantage of using Convolutional Neural Networks (CNNs) in image processing tasks?
✓They can capture spatial hierarchies in data through local connections and shared weights.
CNNs are effective in image processing because they capture spatial hierarchies by using convolutional layers with local receptive fields and weight sharing, which detect local features like edges and build up to more complex features.
✗They require less data to train.
Weight sharing does make a CNN far more parameter-efficient than a fully connected network on images, so there is something to this. But deep CNNs remain notoriously data-hungry, which is why transfer learning from large pre-trained backbones is standard practice.
✗They are less computationally intensive than other models.
Convolutions over high-resolution images are expensive, and CNNs are typically what GPUs were bought for. The saving is in parameter count, not in raw computation.
✗They do not require activation functions.
CNNs need non-linearity just as much as any other network. ReLU after each convolution is the near-universal default.
5. Which type of neural network is specifically designed to handle sequential data by maintaining a hidden state that captures information about previous inputs?
✗Feedforward Neural Network
Information flows one way with no cycles, so nothing persists between inputs. Each example is processed with no knowledge of what came before.
✗Convolutional Neural Network
Detects local patterns with filters over a fixed window. A 1D CNN can be applied to sequences, but its context is limited to the receptive field rather than an accumulating hidden state.
✓Recurrent Neural Network
Recurrent Neural Networks (RNNs) are designed for sequential data by having connections that form directed cycles, allowing information to persist in the hidden state across sequence steps.
✗Autoencoder
Describes a training objective — compress then reconstruct — rather than a way of handling sequences. An autoencoder can be built from recurrent layers, but nothing about the architecture is inherently sequential.
6. Which optimization technique involves adjusting the learning rate during training based on the moving average of past gradients?
✓Adam Optimizer
The Adam Optimizer combines the advantages of AdaGrad and RMSProp, using estimates of first and second moments of gradients to adaptively adjust learning rates during training.
✗Stochastic Gradient Descent
Uses only the current mini-batch gradient with a single fixed learning rate. It maintains no history at all.
✗Momentum
Does keep a moving average of past gradients, so half of this is right. What it does not do is adapt the learning rate per parameter — Adam is momentum plus adaptive scaling.
✗Adaptive Gradient Algorithm (AdaGrad)
Adapts the rate per parameter, but from the cumulative sum of all past squared gradients, with no momentum term and no decay. The sum only grows, so the effective learning rate falls monotonically until progress stops.
7. What is the purpose of using 'dropout' in deep learning models?
✓To prevent overfitting by randomly omitting neurons during training
Dropout is a regularization technique that helps prevent overfitting by randomly dropping neurons during training, which forces the network to learn more robust features.
✗To increase model complexity
Dropout reduces the effective capacity available on any given training step. That constraint is the mechanism by which it regularises.
✗To speed up training by reducing computation
No parameters are removed and all weights still receive updates across batches. Dropout usually needs more epochs to converge, not fewer.
✗To ensure all neurons are equally important
It does discourage over-reliance on any single unit, so redundancy is a genuine side effect. But the objective is generalisation, not equality of importance.
8. Which of the following is a popular open-source deep learning framework developed by Google Brain?
✗Theano
One of the earliest deep learning libraries, developed at MILA, University of Montreal. Major development ceased in 2017.
✓TensorFlow
TensorFlow is an open-source deep learning framework developed by the Google Brain team, widely used for building and deploying machine learning models.
✗PyTorch
Created by Meta's AI research group, not Google. It is TensorFlow's main rival rather than its sibling.
✗Caffe
Developed at UC Berkeley by the Berkeley Vision and Learning Center, originally for computer vision work.
22 more questions in the app
Practise the full 30-question set with a timer, scoring and progress tracking.