Fundamentals of AI and ML

A trained fraud-detection model is deployed to a production API. When the API receives a new transaction, the model produces a fraud probability score in real time. What is the term for this production-time step where the model generates a prediction from new input?

• ATraining, where the model updates its internal weights using labelled examples and a loss function.
• BFine-tuning, where a pre-trained model is further adapted on a domain-specific labelled dataset.
• CEvaluation, where model performance is measured against a held-out test set using metrics such as AUC-ROC.
• DInference, where a deployed model applies learned parameters to new input data to produce an output.check_circle Correct

Why A is wrong: Tempting because training is the most prominent phase of the ML lifecycle, but training occurs before deployment. The stem describes a post-deployment step on new, unseen data.

Why B is wrong: Tempting because fine-tuning is a common deployment-adjacent concept, but it is still a training-time activity. It does not describe the act of generating a prediction from a deployed model on live input.

Why C is wrong: Tempting because evaluation also uses a model against real data, but it is a development-time step that measures quality. The stem describes live production scoring, not a benchmarking activity.

Why D is correct: Inference is the ML lifecycle phase in which a trained model uses its fixed parameters to process new inputs and generate predictions or outputs. This is exactly what happens when the API scores a live transaction.

A logistics company trains a demand forecasting model and then deploys it as a REST API endpoint behind a load balancer. Six months later, a newly hired engineer rewrites the data pre-processing pipeline, causing feature values to be computed differently from how they were computed during training. The model begins producing erratic forecasts. This situation exemplifies which ML lifecycle challenge?

• ATraining-serving skew, where the features supplied to the model during inference differ from those seen during training due to inconsistent pre-processing.check_circle Correct
• BConcept drift, where the relationship between demand signals and actual order volumes has changed in the market since the model was trained.
• CModel decay, where the statistical patterns learned during training become less relevant as new products and routes are introduced over time.
• DOverfitting, where the model memorised noise in the original training data and now fails to generalise to slightly different feature values.

Why A is correct: Correct. Training-serving skew occurs when the transformation applied to raw inputs at inference time diverges from the transformation applied during training. Even though the underlying raw data is similar, the model receives feature values outside its training distribution, producing unreliable outputs. This is a recurring operational challenge in the deployment and monitoring phase of the ML lifecycle.

Why B is wrong: Tempting because concept drift also causes previously accurate models to produce erratic forecasts, and it is a legitimate post-deployment risk. However, concept drift is a gradual market or environmental change, whereas the described problem began immediately after a specific code change. The direct causal link to the pipeline rewrite points to training-serving skew rather than an external distribution shift.

Why C is wrong: Tempting because model decay describes a temporal degradation that is common in logistics forecasting as the product catalogue and route network evolves. However, model decay is a slow process driven by real-world changes, not an immediate effect triggered by a code change in the pre-processing pipeline.

Why D is wrong: Tempting because overfitting does cause sensitivity to input variation, but overfitting is a training-phase problem that would have been visible in validation metrics before deployment. The scenario describes a model that performed well for six months before a specific pipeline change, which rules out pre-existing overfitting as the explanation.

A startup is deciding whether to use machine learning or a deterministic algorithm to calculate the shortest delivery route between two fixed points on a well-mapped city grid. Which consideration most strongly suggests ML is NOT necessary here?

• AThe city grid has many intersections, which increases computational complexity for any approach.
• BThe startup lacks GPU infrastructure required to run machine learning models in production.
• CThe problem has a mathematically optimal solution computable by a deterministic algorithm such as Dijkstra's, with no need to learn from data.check_circle Correct
• DThe dataset of past delivery routes is too small to train a reliable model.

Why A is wrong: Computational complexity is a concern for all approaches but does not determine whether ML is appropriate. Established graph search algorithms handle large city grids efficiently without ML.

Why B is wrong: Infrastructure limitations are a deployment concern, not a reason that ML is conceptually inappropriate. Many ML models run on CPUs. The real reason here is that the problem has an exact classical solution.

Why C is correct: When a problem has a known, exact, deterministic solution, introducing ML adds complexity without benefit. Shortest-path problems on a graph are solved optimally by classical algorithms. ML is not needed and would likely produce inferior results compared to an exact method.

Why D is wrong: Data scarcity is one factor in ML feasibility, but the deeper reason to avoid ML here is that the problem does not require learning from data at all - it has an exact mathematical solution. Framing it as a data problem misses the core point.

A data science team labels thousands of customer reviews as positive, neutral, or negative, then trains a model on those labels. Which category of machine learning does this training approach represent?

• ASupervised learning, where labelled training examples teach the model to map inputs to correct output categories.check_circle Correct
• BUnsupervised learning, where the model groups similar reviews without any provided labels.
• CReinforcement learning, where an agent receives reward signals based on actions taken in an environment.
• DSelf-supervised learning, where the model creates its own training signal from the structure of unlabelled text.

Why A is correct: Supervised learning trains on input-label pairs. Providing labelled sentiment categories (positive, neutral, negative) and using them to train a classifier is the canonical supervised learning workflow.

Why B is wrong: Tempting because unsupervised learning also processes text data, but it works without labels. The stem explicitly states that labels were created and used, ruling this out.

Why C is wrong: Tempting because reinforcement learning is a prominent ML paradigm, but it requires an agent-environment loop with reward signals, not a static labelled dataset of reviews.

Why D is wrong: Tempting because large language models do use self-supervised pre-training on text, but that approach does not involve human-created sentiment labels. The stem describes explicit human annotation followed by training.

An ML engineer sets up automated alerts that fire when the Kullback-Leibler divergence between the live prediction score distribution and the baseline training distribution exceeds a defined threshold. This practice belongs to which phase of the ML lifecycle?

• AData labelling, where annotators flag ambiguous examples that fall near the decision boundary.
• BModel selection, where multiple candidate architectures are compared using divergence metrics computed on the held-out validation set.
• CFeature engineering, where distributional properties of raw features are measured to guide transformation choices.
• DModel monitoring, where production behaviour is continuously compared against an established baseline to detect drift or degradation.check_circle Correct

Why A is wrong: Tempting because boundary proximity is related to the divergence concept, but data labelling is a pre-training activity performed on offline datasets. KL divergence monitoring of live prediction distributions is a production-time activity that cannot occur during labelling.

Why B is wrong: Tempting because KL divergence is a legitimate model comparison tool, but model selection uses held-out data before deployment. The scenario explicitly describes live prediction distributions, confirming the model is already in production, so this cannot be model selection.

Why C is wrong: Tempting because feature engineering does involve analysing distributions, but that analysis operates on offline training data before model training. The described monitoring uses live prediction scores from a deployed model, placing it firmly in the post-deployment monitoring phase rather than pre-training feature engineering.

Why D is correct: Correct. Model monitoring is the post-deployment phase responsible for tracking production model health. Measuring KL divergence between live score distributions and a training baseline is a canonical technique for detecting prediction drift, which often precedes measurable accuracy loss. Automated thresholded alerts are standard monitoring infrastructure.