8 of the 56 questions in this set, with the correct answer marked and every option explained.
1. Production grade machine learning challenges are addressed by implementing an important concept:
✗Tensorflow Extended (TFX)
One framework for building ML pipelines. A specific implementation, where the question asks for the underlying concept it implements.
✗Directed Acyclic Graphs (DAGs)
The structural form a pipeline takes — steps ordered with no cycles. It describes the shape rather than the practice.
✗Orchestrators
The tools that schedule and run pipeline steps, such as Airflow or Kubeflow Pipelines. Execution machinery rather than the concept itself.
✓Machine learning pipelines
ML pipelines provide support for automating, monitoring and maintaining a model as you continue to train it over its lifetime.
2. MLOps applies established software engineering practices to machine learning systems. Which of the following are practices carried over from modern software development? (Select THREE)
✓Testability
Code written so it can be tested automatically — unit tests, integration tests, and in ML additionally tests over data and model behaviour. A core software engineering property that MLOps inherits directly.
✓Monitoring
Observing a running system in production. Software engineering monitors latency, errors and saturation; MLOps adds prediction quality, data drift and training-serving skew on top of the same discipline.
✓Version control
Tracking changes so any state can be reproduced and reviewed. ML extends it beyond source code to datasets, features and model artefacts, but the practice comes straight from software development.
✗Fast model training
A machine learning concern about compute efficiency and iteration speed. It has no counterpart in general software engineering, which is what the question asks about.
3. More often than not, ML systems can fail the users it serves. In this context, what is representational harm?
✗Making predictions and decisions that preclude certain groups from accessing resources or opportunities.
Allocative harm — being denied credit, a job or a service. It concerns material outcomes rather than how a group is depicted.
✗Giving skewed outputs more frequently for certain groups of users
Quality-of-service harm, where the system simply works worse for some people. Real, but distinct from reinforcing a stereotype.
✗Inferring prejudicial links between certain demographic traits and user behaviors.
Describes the flawed inference inside the model. Frequently the mechanism behind representational harm, but the harm itself is the resulting portrayal.
✓The amplification or negative reflection of certain groups stereotypes.
Representative harms occur when systems reinforce the subordination of some groups along the lines of identity. The amplification or negative reflection of certain groups stereotypes. This is a prototypical way an ML system may fail the users it serves.
4. Accurate labels are necessary to properly train supervised models. Many times, human subjects known as raters perform this labeling effort. What are the main categories of human raters? (check all that apply).
✓Subject matter experts
Generalists usually come from crowdsourcing sites. SME or subject matter experts : classical example is radiologists labeling medical images for automated diagnosis tools. Users can provide labels within your application. A classical example is photo tagging.
✓Your users
✗Aggregators
Aggregation is what you do with multiple raters' labels to settle disagreement — majority vote, weighted consensus. A method, not a type of rater.
✗Generalists
Non-specialist crowd workers are widely used in practice, so this has some claim. In this framing the categories are subject matter experts, who supply domain knowledge, and your own users, who label implicitly through their behaviour.
5. A cardiologist labeling MRI images is a typical example of Direct Labeling.
✓False
MRI labeling by a medical specialist is an example of Human Labeling and not Direct Labeling.
✗True
Direct labelling means the label emerges automatically from how the system is used — a click, a purchase, a skip. A cardiologist annotating scans is deliberate expert labelling, which is slow and costly by comparison.
6. What formula represents a dataset shift?
✗Ptrain(x|y)≠Pserve(y|x) and Ptrain(y|x)≠Pserve(y|x)
Not well formed — it compares P(x|y) against P(y|x), which are different conditional distributions and cannot be set against each other.
✗Ptrain(y|x)≠Pserve(y|x) and Ptrain(x)=Pserve(x)
This is concept shift, the case where the input-to-label relationship changes while inputs stay put. A subtype rather than the general definition.
✗Ptrain(y|x)=Pserve(y|x) and Ptrain(x)≠Pserve(x)
This is covariate shift — inputs move, relationship holds. Also a subtype.
✓Ptrain(y|x)≠Pserve(y|x)
The most generic case of distribution skews is when the joint distribution of inputs and outputs differs between training and serving.
7. What formula represents a concept shift?
✗Ptrain(y|x)=Pserve(y|x) and Ptrain(x)≠Pserve(x)
Covariate shift, the mirror image. The inputs have drifted but what they imply about the label has not changed.
✗Ptrain(x|y)≠Pserve(y|x) and Ptrain(y|x)≠Pserve(y|x)
Malformed for the same reason as before — two different conditionals placed on either side of one comparison.
✗Ptrain(y|x)≠Pserve(y|x)
The umbrella definition of dataset shift. Concept shift additionally requires the input distribution to be unchanged.
✓Ptrain(y|x)≠Pserve(y|x) and Ptrain(x)=Pserve(x)
When the joint distribution of inputs and outputs differs between training and serving but the probability distributions of the features (x) is matching between train and serving then we say concept ('y') is drifted.
8. What measure is typically used to determine the degree of data drift?
✓Chebyshev distance (L-infinity)
Chebyshev distance is defined as max(|xi−yi|)
✗Manhattan distance (L1)
Adds absolute differences across every dimension. A valid metric, but summing means a single sharply shifted bucket gets diluted among the rest.
✗Hamming distance
Counts positions where two equal-length strings differ. Built for categorical or binary sequences, not continuous distributions.
✗Euclidean distance (L2)
Straight-line distance, and a common default. Like L1 it aggregates across dimensions, whereas L-infinity surfaces the single worst deviation — which is what should trigger a drift alert.
48 more questions in the app
Practise the full 56-question set with a timer, scoring and progress tracking.
