IT Passport Examination (IP) | Technology: AI, Machine Learning, and Generative AI Questions 08

Each customer record pairs its features with the correct outcome label, "churned" or "remained." Learning that relationship to predict a new customer's outcome is supervised learning. Unsupervised learning instead seeks structures or groups without supplied answer labels.

No correct customer category is supplied beforehand; the goal is to group customers with similar purchase patterns. This is an appropriate use of unsupervised learning. Supervised learning would instead train a classifier from known category labels.

The robot tries actions in an environment and improves its policy based on the resulting positive and negative rewards, which is reinforcement learning. It is not directly given the correct route for every situation as a label, as it would be in supervised learning.

The model fits its training data extremely well but fails to generalize to unseen data, so this is overfitting. If the model were untrained or substantially undertrained, its performance on the training data would not normally be this high either.

Final generalization performance should be measured on test data kept separate from both training and parameter tuning. Reusing training data or data repeatedly consulted during tuning measures adaptation to that data and can overstate performance on genuinely unseen customers.

A foundation model is pretrained on broad, large-scale data and can be adapted to multiple downstream tasks, such as question answering or image recognition, through methods such as fine-tuning. This differs from a system restricted to a single fixed-rule process.

A hallucination occurs when generative AI produces false information in a plausible form. A deepfake is a realistic synthetic or manipulated representation such as a person's video or voice, so it is distinct from inventing factual case information in this scenario.

Human-in-the-loop operation incorporates human review or judgment into AI processing, so an officer should examine the rationale and make the recorded final decision. Automatically accepting AI decisions removes human involvement, while never using the output does not create human-AI collaboration.

Because the historical data itself is concentrated in one group, the organization should inspect its composition, evaluate outcomes by group and remediate the data or model where needed. Removing one explicit attribute may not eliminate bias carried by correlated variables such as location, and an overall accuracy figure can hide harm to particular groups.

Prompt engineering designs the instructions, context, examples and output format supplied to a model and improves them iteratively based on evaluation, without retraining the model itself. Transfer learning and fine-tuning instead reuse learned knowledge while performing additional training on the model side.