AI Concepts and Techniques practice questions

Precision

Why it fails: Precision only considers the correctness of positive predictions, ignoring false negatives.

Recall

Why it fails: Recall only measures how many actual positives are captured, ignoring false positives.

Accuracy

Why it fails: Accuracy can be high even if the model predicts 'no default' for all cases, which fails to capture the minority class.

Reinforcement learning

Why it fails: Reinforcement learning is for sequential decision-making in dynamic environments, not static recommendation.

Unsupervised clustering

Why it fails: Clustering can segment users but does not directly predict which product a user will buy.

Self-supervised learning

Why it fails: Self-supervised learning is typically used for pre-training representations, not directly for recommendation.

Language translation

Why it fails: Translation changes language but does not extract phrases.

Text classification

Why it fails: Text classification assigns categories but does not extract phrases.

Sentiment analysis

Why it fails: Sentiment analysis determines emotional tone, not key phrases.

L1 regularisation (Lasso)

Why it fails: L1 can zero out coefficients, which may be too aggressive if all features are relevant; L2 is generally safer for reducing variance without losing features.

Dropout

Why it fails: Dropout is used in neural networks, not linear regression.

Data augmentation

Why it fails: Data augmentation increases training data, but the question asks for a regularisation technique.

The model should use a different evaluation metric like BLEU

Why it fails: BLEU scores n-gram overlap for generated text such as translation or summarisation, so it cannot evaluate a binary sentiment classifier. Switching to BLEU is correct when the task produces free-form generated text and you need to measure similarity against reference outputs.

The model is excellent because accuracy is high

Why it fails: High accuracy here merely reflects class imbalance, since the majority-class baseline already reaches 0.85; the 0.92 figure hides weak precision (0.75) and recall (0.80). Accuracy is genuinely useful when classes are balanced and misclassification costs are equal, but this scenario demands F1 or per-class metrics instead.

The model has high variance and is overfitting

Why it fails: Overfitting and high variance describe a large gap between training and test performance; the stem supplies only test-set scores, with no training metrics to compare. Variance analysis becomes the justified conclusion when training accuracy substantially exceeds validation accuracy for the same model.

Support vector machine (SVM) with RBF kernel

Why it fails: SVMs do not natively provide feature importance and require feature scaling.

Logistic regression

Why it fails: Logistic regression coefficients can indicate direction but not direct importance, and requires encoding categorical variables.

k-nearest neighbours (k-NN)

Why it fails: k-NN is sensitive to scaling and provides no feature importance.

Use a smaller batch size

Why it fails: Larger batches reduce gradient variance, so smaller batches would add noise and worsen the fluctuation. Smaller batches suit memory-constrained fine-tuning or regularisation via stochasticity, but here the instability stems from noisy gradient estimates that only a larger batch or lower learning rate would calm.

Increase the learning rate

Why it fails: Raising the learning rate enlarges each weight update, amplifying the epoch-to-epoch oscillation rather than damping it. A larger rate suits speeding convergence on a stable, well-conditioned loss landscape, not a validation curve already swinging wildly.

Increase the number of epochs

Why it fails: Increasing epochs does not address the epoch-to-epoch variance; it merely extends training, so the same instability persists or worsens. It is tempting because extra epochs help when a model is underfit and still improving, but here the fluctuation points to optimisation dynamics, not insufficient training time.

Use self-supervised learning to pretrain on the unlabeled data, then fine-tune on the labeled data

Why it fails: Self-supervised learning is a valid approach, but semi-supervised learning is more direct when you have some labels and many unlabeled examples.

Train a supervised classifier on only the 10,000 labeled reviews

Why it fails: Ignoring the unlabeled data wastes valuable information that could improve performance.

Use reinforcement learning with the unlabeled data as rewards

Why it fails: Reinforcement learning requires a reward signal from the environment, which is not available in this static dataset.

The batch size is too small

Why it fails: Batch size affects gradient noise but would not directly cause this perplexity-BLEU divergence pattern.

The model is underfitting the training data

Why it fails: Underfitting would cause high perplexity and low BLEU; here perplexity is low but BLEU dropped.

The learning rate is too high

Why it fails: A high learning rate would typically cause instability, not a drop in perplexity.

Large language model (LLM)

Why it fails: LLMs are optimized for text generation, not image synthesis.

Variational autoencoder (VAE)

Why it fails: Variational autoencoders (VAEs) optimise a variational lower bound on the log-likelihood, which forces a trade-off between reconstruction fidelity and latent-space regularity, often producing blurry outputs that lack the fine-grained detail required for realistic product images. This approach is tempting because VAEs excel at learning smooth, continuous latent representations for tasks like anomaly detection or controlled interpolation, where generating slightly varied versions of existing designs is the goal.

Diffusion model

Why it fails: Diffusion models are also effective for image generation, but GANs are more established for this specific task; however, both are plausible. But GAN is the classic answer.

Increase the number of epochs

Why it fails: More epochs can lead to overfitting without other regularization.

Use L2 regularization

Why it fails: L2 regularization helps but is generally less effective for CNNs compared to dropout and augmentation.

Add more convolutional layers

Why it fails: Adding more layers increases model capacity, which can worsen overfitting.

BLEU score

Why it fails: BLEU is for evaluating machine translation outputs.

Perplexity

Why it fails: Perplexity is used for language models, not binary classifiers.

Accuracy

Why it fails: Accuracy can be misleading with imbalanced classes; not always the best metric.

Train a custom model from scratch on the policy documents each month

Why it fails: Training from scratch each month is prohibitively expensive and slow, and it cannot ingest monthly document changes without full retraining — exactly the cost the team rules out. It is tempting because bespoke training gives maximum control over domain vocabulary, and it would suit a stable corpus with abundant labelled data and no update cadence.

Use a larger foundation model with a longer context window and paste all documents into each prompt

Why it fails: Pasting every policy document into each prompt overflows the context window and inflates token cost and latency per request; retrieval of only relevant passages is required. It is tempting because long-context prompting needs no indexing pipeline, and it would be correct for a small, static document set that fits comfortably.

Fine-tune a base LLM on the policy documents monthly

Why it fails: Fine-tuning bakes knowledge into weights, so monthly document changes still demand retraining runs — the expense the team explicitly cannot afford. It is tempting because fine-tuning reliably shapes tone and output format, and it would be correct when the goal is behavioural style rather than frequently changing factual content.

Accuracy

Why it fails: Accuracy would be 99.9% by predicting all legitimate — misleading.

Recall

Why it fails: Recall is important but does not capture false positives.

Precision

Why it fails: Precision is important but does not capture recall trade-off.

Set the temperature to 0

Why it fails: Temperature affects output randomness, not injection vulnerability.

Use a system prompt that instructs the model to ignore injection attempts

Why it fails: System prompts can be overridden by user input; not reliable.

Use a larger, more powerful LLM

Why it fails: Larger models may still be vulnerable to injection.

Remove dropout layers to allow more gradient flow

Why it fails: Removing dropout may increase overfitting but not necessarily help convergence.

Use a smaller batch size to reduce memory

Why it fails: Smaller batch size introduces noise, which may not help convergence.

Increase the learning rate by 10x

Why it fails: A high learning rate may cause divergence, not convergence.

Transfer learning

Why it fails: Transfer learning reuses a model trained on one task or domain and adapts it to another. Although the pipeline involves two stages, the first stage is unsupervised learning on the same unlabeled vitals, not a separate source task. The defining characteristic here is the combination of unlabeled and labeled data, which is semi-supervised, not transfer learning.

Reinforcement learning

Why it fails: Reinforcement learning trains an agent through reward signals from interactions with an environment, not through labeled examples or unsupervised feature extraction. The scenario describes learning from unlabeled vitals and a small labeled set, with no reward mechanism or sequential decision-making agent, so reinforcement learning does not describe this training pipeline.

Self-supervised learning

Why it fails: Self-supervised learning generates supervisory signals from the data itself, such as predicting masked values, without any external labels. While the initial unsupervised phase may resemble self-supervision, the pipeline explicitly includes a fine-tuning stage with a small labeled set, so the complete paradigm is not purely self-supervised and this choice mischaracterizes the end-to-end training.

Set top-p to 0.1

Why it fails: Top-p 0.1 restricts sampling to the smallest token set whose cumulative probability reaches 10%, sharply curtailing creative variation and often producing repetitive, generic replies. Top-p governs nucleus sampling breadth; a moderate value, such as 0.9, preserves creativity while trimming the long tail of unlikely tokens.

Increase top-k to 100

Why it fails: Raising top-k to 100 widens the sampling pool to the 100 most probable tokens, admitting lower-probability, off-topic tokens and increasing nonsensical output. Top-k is intended to constrain sampling breadth; a smaller value, such as 10, would restrict candidates and reduce off-topic responses.

Increase temperature to 0.9

Why it fails: Temperature 0.9 flattens the probability distribution, raising the chance of low-probability tokens and thus more off-topic or nonsensical output. Temperature controls randomness; a lower value, such as 0.3, sharpens the distribution and is the correct adjustment when creativity must be constrained.

Supervised classification with logistic regression

Why it fails: Logistic regression needs labelled examples of both classes to learn a decision boundary, and attacks are too scarce to train it. It is tempting because supervised classification is the standard route when labelled data exists, but here labels are insufficient, so anomaly detection on normal traffic fits.

Reinforcement learning

Why it fails: Reinforcement learning learns through reward signals from environment interaction, requiring a simulation or feedback loop that this static traffic dataset cannot supply. It is tempting because it excels at sequential decision tasks such as adaptive intrusion response, but anomaly detection on predominantly unlabelled data calls for unsupervised methods.

Semi-supervised learning

Why it fails: Semi-supervised learning still requires a meaningful pool of labelled attack samples to propagate labels; with very few, it cannot reliably characterise breaches. It is tempting because it exploits abundant unlabelled data, but the scenario's near-absence of attack labels points to unsupervised anomaly detection instead.

Use early stopping without any other regularisation

Why it fails: Early stopping halts training when validation loss rises but adds no penalty to weights, so 1,000 images still overfit. It tempts because it is cheap and effective alongside other methods, and would be correct as one component of a combined regularisation strategy.

L1 regularisation and batch normalisation

Why it fails: L1 drives weights to zero for sparsity, and batch normalisation stabilises activations; neither directly constrains model capacity the way dropout or weight decay do. It tempts because both are legitimate regularisers, and would be correct for feature selection or faster convergence.

Increase learning rate and use momentum

Why it fails: A higher learning rate and momentum accelerate optimisation but do not penalise complexity, so they can worsen overfitting on 1,000 images. It tempts because both speed convergence, and would be correct when training is slow or stuck in poor minima.