8 of the 30 questions in this set, with the correct answer marked and every option explained.
1. What is the primary purpose of using residual connections (skip connections) in deep convolutional neural networks like ResNet?
✗To increase the depth of the network without encountering vanishing gradients.
Very nearly the same claim, and true — skip connections give gradients a short path back and are what made hundred-layer networks trainable. The keyed answer states the mechanism that produces this benefit: making it trivial for a block to represent the identity function, so extra depth can never hurt.
✗To enforce sparsity in the network weights, improving generalization.
Sparsity comes from L1 regularisation or explicit pruning. Residual connections add a path, they do not zero out weights.
✓To allow the network to learn identity mappings, facilitating training of deeper networks.
Residual connections enable the network to learn identity mappings by adding the input of a layer to its output. This helps in training very deep networks by mitigating the vanishing gradient problem, allowing gradients to flow directly through the skip connections.
✗To reduce the computational complexity by bypassing certain layers.
The skip path is added to the layer's output, not used instead of it. Both branches are computed, so cost goes slightly up rather than down.
2. Which of the following best describes the main advantage of the Adam optimization algorithm over traditional stochastic gradient descent in training neural networks?
✗Adam ensures convergence to the global minimum in non-convex optimization problems.
No optimiser can guarantee this on a non-convex surface. Adam typically converges faster, with no promise about which minimum it reaches.
✗Adam uses second-order derivatives to accelerate convergence.
A common confusion. Adam tracks second moments of the gradient — its uncentred variance — not second derivatives of the loss. It never computes a Hessian.
✓Adam uses per-parameter learning rates, adapting the updates for each parameter individually.
Adam (Adaptive Moment Estimation) computes adaptive learning rates for each parameter by keeping track of first and second moments of gradients, allowing for efficient training of neural networks with sparse gradients.
✗Adam introduces momentum to smooth out updates.
True, and part of what Adam does through its first-moment estimate. But momentum alone is not new — it predates Adam. The distinguishing feature is combining momentum with per-parameter adaptive rates.
3. In Generative Adversarial Networks (GANs), what is the role of the discriminator?
✗To minimize the loss function of the generator.
The discriminator works against the generator. Its objective is adversarial — it tries to maximise its own accuracy, which raises the generator's loss.
✓To distinguish between real and generated data samples.
In a GAN, the discriminator's role is to classify inputs as real (from the training data) or fake (generated by the generator). The discriminator and generator are trained simultaneously in a minimax game, where the generator tries to fool the discriminator.
✗To enforce a prior distribution on the latent space.
That is the KL divergence term in a VAE. A standard GAN places no explicit prior constraint on its latent space.
✗To generate new data samples resembling the training data.
The generator's job. The discriminator only ever judges.
4. What is the 'credit assignment problem' in reinforcement learning?
✗Allocating computational resources among multiple agents.
A distributed systems concern. Credit assignment is about attributing outcomes to actions, not about scheduling compute.
✗Balancing exploration and exploitation during training.
The exploration-exploitation dilemma — a separate central problem in reinforcement learning, concerning whether to try something new or exploit what already works.
✗Distributing rewards among agents in multi-agent systems.
Multi-agent credit assignment, a related but narrower variant. The classic problem exists with a single agent: which of the hundred moves before checkmate deserves the credit?
✓Determining which action is responsible for a delayed reward.
The credit assignment problem involves figuring out which actions taken by an agent led to a particular reward, especially when the reward is delayed. This is critical for updating policies that maximize cumulative rewards.
5. Which of the following methods is used to explain individual predictions of a machine learning model by approximating it locally with an interpretable model?
✗Grad-CAM (Gradient-weighted Class Activation Mapping)
Produces a local explanation too, but through gradients flowing into convolutional feature maps to highlight image regions. It is specific to CNNs rather than model-agnostic, and it fits no surrogate model.
✗SHAP (SHapley Additive exPlanations)
Also explains individual predictions, which makes this close. The mechanism differs: SHAP computes Shapley values from cooperative game theory, whereas LIME fits an interpretable surrogate on perturbed samples around the instance.
✗PCA (Principal Component Analysis)
A dimensionality reduction technique. It finds directions of maximum variance in data and explains no model's prediction.
LIME approximates the model locally around the prediction with a simpler interpretable model (like linear regression), to explain why the model made a certain prediction.
6. Which of the following best describes 'algorithmic bias' in machine learning models?
✓A systematic error introduced by the algorithms favoring certain outcomes.
Algorithmic bias refers to systematic errors in AI systems that result in unfair outcomes, such as privileging one group over others, often due to biased training data or flawed algorithms.
✗Random noise affecting the predictions.
Noise is random and unsystematic by definition. Bias is precisely the opposite — a consistent lean in one direction.
✗The tendency of models to overfit to the training data.
Overfitting is a generalisation failure. It can amplify existing bias, but a perfectly generalising model can still be systematically unfair.
✗An error due to insufficient model capacity.
That is underfitting, sometimes called bias in the bias-variance sense. Worth separating the two meanings: statistical bias in that trade-off is a different concept from algorithmic bias against groups of people.
7. What is the main innovation of the Transformer architecture compared to traditional RNN-based sequence models in NLP?
✗It incorporates reinforcement learning for sequence generation.
Transformers are trained with supervised or self-supervised objectives. RLHF is applied to some models much later and is not architectural.
✓It relies entirely on attention mechanisms, dispensing with recurrence.
The Transformer architecture uses self-attention mechanisms to process sequences, allowing for parallelization and better handling of long-range dependencies, unlike RNNs which process sequentially.
✗It uses autoencoders for dimensionality reduction.
No autoencoder is involved. The encoder-decoder structure of a transformer serves sequence-to-sequence mapping, not compression.
✗It uses convolutional layers to process sequences.
Convolutional sequence models already existed before the transformer. Its break with the past was using neither convolution nor recurrence.
8. In object detection, what is the purpose of using anchor boxes in algorithms like Faster R-CNN?
✗To normalize the input images for consistent scaling.
Image normalisation is a preprocessing step. Anchor boxes operate inside the detection head, on candidate regions rather than pixels.
✓To provide predefined bounding box shapes for predicting object locations.
Anchor boxes are predefined bounding boxes of various scales and aspect ratios used as references for predicting object locations and sizes in object detection models like Faster R-CNN.
✗To augment the dataset with synthetic examples.
Augmentation generates new training images. Anchors are fixed reference shapes used at prediction time, adding no data.
✗To reduce computational complexity by limiting the search space.
A genuine side effect — anchors avoid an exhaustive sliding-window search. But their purpose is to give the network sensible priors of varying scale and aspect ratio, which it then refines through offset regression.
22 more questions in the app
Practise the full 30-question set with a timer, scoring and progress tracking.