AI Evaluation and Fairness Flashcards

Flashcards covering key vocabulary related to AI evaluation and fairness.

Fairness

Treating all individuals and groups equitably, without favoring one group over another.

Bias

Producing systematically prejudiced outcomes.

Traceability

Ability to track and document the decision-making processes, data sources, and changes made to an AI model throughout its lifecycle.

Explainability

Ability of an AI system to provide understandable reasons or explanations for its outputs or decisions.

Responsibility

Assigning accountability for the actions and outcomes of AI systems, entails mitigating risks such as harm or misuse.

Liability

Legal responsibility assigned to individuals or organizations for the actions, outcomes, or consequences of an AI system.

Gender Bias

NLP models associate certain professions or behaviors with specific genders.

Racial Bias

NLP models underperform on languages or dialects spoken by minority groups.

Cultural Bias

NLP chatbot fails to understand cultural nuances.

Data Transparency

Knowing the sources of data used to train NLP models.

Versioning

Keeping track of changes in an NLP model.

Reproducibility

Allows developers to reproduce results and check for consistency over time.

Explainability Benefit

Helps users trust NLP systems by explaining outputs.

Content Moderation

NLP models used in social media platforms need to avoid spreading misinformation or hate speech.

Ethical Model Development

Developers must ensure that datasets used for training NLP models are ethically sourced and do not contain biases.

Corrective Mechanisms

Implementing processes to fix or update models when biased or unethical behavior is identified.

Adversarial Examples

Inputs that have been deliberately perturbed to mislead a model.

Synonym Substitution

Changing words to synonyms to confuse a model.

Character-level Attacks

Introducing typos to confuse a model.

Paraphrasing

Rephrasing text to confuse a model.

Clean Data

Remove duplicates, irrelevant content, or highly noisy data.

Cohen's kappa agreement

Take into account the agreement by chance.

Baseline

A simple model or method used as a point of comparison to evaluate the performance of more advanced models.

Random Baseline

Assigns labels randomly.

Majority Class Baseline

Always predicts the most frequent class in the dataset.

BLEU (Bilingual Evaluation Understudy)

Calculates the overlap of n-grams between the machine-generated translation and the reference translation.

ROUGE-N

Measures overlap of n-grams.

ROUGE-L

Measures the longest common subsequence (LCS).

MMLU (Massive Multitask Language Understanding)

Tests knowledge and reasoning across 57 academic subjects.

HellaSwag

Tests commonsense reasoning and narrative plausibility.

TruthfulQA

Tests truthfulness, especially in response to tricky or adversarial questions.

BIG-bench (Beyond the Imitation Game)

A giant collection of 200+ tasks to probe creativity, reasoning, morality.

HLE (Humanity’s Last Exam)

A new gold-standard benchmark to test whether AI systems have reached expert human reasoning.