Knowledge check

Quiz

Complete

Evaluate your understanding of this unit by completing the Knowledge Check.

A healthcare company develops an AI to predict patient readmission rates. The initial model shows high accuracy overall, but when analysed across different age groups, it significantly underestimates the readmission risk for elderly patients. What is the most likely underlying cause of this disparity?

A) High variance in the overall dataset.B) Representation bias in the training data, with fewer elderly patient records or less detailed information for this group.C) The model algorithm inherently favours younger patient data.D) Overfitting to the majority age groups in the training data.

A company uses an AI-powered tool to score job applications. After deployment, it's observed that candidates from less prestigious universities consistently receive lower scores, even when their skills and experience are comparable to graduates from top-tier institutions. Which bias is most likely at play here, and what could have been a contributing factor during data collection?

A) Measurement bias, potentially due to using university prestige as a direct feature without considering individual achievements.B) Historical bias, reflecting past hiring trends that favoured certain universities.C) Representation bias, where the training data lacked sufficient applications from graduates of less prestigious universities.D) Aggregation bias, where a single scoring model is applied without considering the diverse backgrounds of applicants.

When developing a facial recognition system, developers train the model primarily on images of one demographic group. Later, the system shows significantly lower accuracy when identifying individuals from other demographic groups. This is a clear example of:

A) High variance leading to poor generalisation.B) Measurement bias in the image capture process.C) Representation bias in the training data.D) Low bias causing overfitting to the dominant group.

A content recommendation system is trained on user interaction data. If users from a certain cultural background tend to interact more with specific types of content, what potential bias could arise, and how might it manifest for new users from that same background?

A) Historical bias, leading to the system recommending content that was popular in the past for that group, even if current preferences have shifted.B) Aggregation bias, where the system fails to recognise the diverse tastes within that cultural group.C) Representation bias, causing the system to overemphasise the preferences of the dominant group in the training data, leading to skewed recommendations for new users from that group.D) Measurement bias, if the user interaction data doesn't accurately reflect their true content preferences.

An AI model is used to predict criminal recidivism. If the training data reflects historical biases in policing and sentencing, leading to certain demographic groups being overrepresented in the "recidivated" category, what is the most critical ethical concern regarding the model's deployment?

A) The model will likely have low overall accuracy.B) The model will perpetuate and potentially amplify existing societal biases, leading to unfair outcomes.C) The model will suffer from high variance due to the inconsistencies in the historical data.D) The model will be difficult to interpret due to the complexity of criminal behaviour.

A company develops a language translation model. If the training data predominantly consists of formal text, how might this bias manifest when users try to translate informal conversations or slang?

A) The model will likely have high variance in its translations.B) The model will likely struggle to accurately translate informal language, potentially producing overly formal or nonsensical outputs.C) This scenario is unlikely to introduce any significant bias.D) The model will likely overfit to the formal language in the training data.

When building a sentiment analysis tool for customer reviews, the training data contains a disproportionately large number of reviews written by users with a high level of technical expertise. How might this representation bias affect the tool's ability to accurately gauge the sentiment of reviews written by less technically savvy customers?

A) The tool might misinterpret the language and tone used by less technical users, potentially misclassifying their sentiment.B) The tool will likely exhibit high variance when analysing reviews from all user groups.C) The tool will likely overfit to the technical language used in the training data.D) The tool's accuracy will remain consistent across all user groups.

A city implements an AI-powered system to optimise traffic flow based on historical traffic data. If the historical data primarily reflects traffic patterns during standard weekday commuting hours, how might this bias affect the system's performance and recommendations during weekends or holidays?

A) The system will likely perform optimally at all times due to the large amount of training data.B) The system might not accurately predict or optimise traffic flow during non-standard hours, as the training data doesn't adequately represent these patterns.C) The system will likely exhibit high variance in its predictions during peak hours.D) The system will overfit to the weekday commuting patterns, leading to inaccurate predictions even during those times.

A company uses an AI model to predict which customers are likely to churn. If the training data disproportionately includes feedback from customers who actively contacted support, how might this affect the model's ability to identify passive churners (customers who leave without explicit complaints)?

A) The model will likely be very accurate in predicting all types of churn.B) The model might be better at identifying active churners but less effective at predicting passive churners, as the training data doesn't fully represent this group.C) The model will likely exhibit high variance in its churn predictions.D) The model will overfit to the characteristics of active churners.

When deploying a machine learning model to classify images, it's noticed that the model performs significantly worse on images taken with older, lower-resolution cameras compared to those from modern high-resolution devices used in the training data. What is the most likely source of this performance disparity?

A) Historical bias in the image dataset.B) Representation bias, where the training data lacked sufficient examples of lower-resolution images.C) Measurement bias due to the different image capture technologies.D) Aggregation bias in how the image features were extracted.

Action item: Knowledge check

In progress