Lecture 15 - Fairness

Other-Race Effect (ORE)

A phenomenon where individuals are better at recognizing faces of their own race compared to those of other races.

Expertise in Face Recognition

Developed through meaningful experiences with faces throughout life; enhanced through exposure to individuals from a homogeneous racial group.

Demographic Influence

The impact that factors like race, gender, and age have on the performance of recognition algorithms.

Performance Discrepancies

Variations in algorithm effectiveness based on different demographic groups.

Threshold Effects

Variations in the performance of algorithms based on differing statistical distributions within sub-populations.

False Match Rate (FMR)

The probability that a biometric system incorrectly matches a biometric sample to an unauthorized user.

Skin Reflectance

The measure of how light is reflected from the skin, which can significantly impact facial recognition performance.

Prototype Learning

A learning method where the recognition system develops templates or prototypes based on input data.

Yoking

Controlling variables in an experiment to ensure demographic comparability between different samples.

Algorithmic Discrimination

Biases that emerge from the misrepresentation or unequal treatment of different demographic groups in algorithm performance.

Casual Conversations Dataset

A collection of videos and annotations used to study facial recognition performance across diverse populations.

Fitzpatrick Scale

A numerical classification schema used to categorize human skin color based on response to ultraviolet light.

Equitable Training

The process of training algorithms using balanced datasets to mitigate biases against underrepresented groups.

Varying Error Rates

Differences in accuracy or performance of algorithms when recognizing faces from various demographic groups.

Image Acquisition Quality

The clarity and quality of images used in training or testing algorithms, impacting their performance.

Neural Network Model

A computational model inspired by the human brain that is used for processing data and learning patterns.

Auto-Associative Learning

A type of learning in neural networks where the system learns by trying to reconstruct its own inputs.

Commercial Off-the-Shelf Algorithms (COTS)

Pre-packaged and widely available algorithms used for various applications, including facial recognition.

False Acceptance Rate (FAR)

The likelihood that a system incorrectly accepts an unauthorized user as an authorized one.

Identification Accuracy

The degree to which a face recognition system correctly identifies individuals.

Homogeneous Population

A group of individuals sharing similar characteristics, often used in studies to reduce variability.

Bias Mitigation

Efforts made to reduce or eliminate discrimination in algorithmic performance across different demographic groups.

Training Dataset Composition

The diversity and characteristics of the data used to train recognition algorithms, influencing their performance.

Performance Evaluation

The assessment of an algorithm’s effectiveness, which can be influenced by demographic factors.

Cross-Race Bias

The tendency of people or systems to show differing levels of accuracy in recognizing faces of different races.

Stereotypes in AI

Preconceived notions or biases embedded in AI models that can lead to unfair outcomes.

Demographic Heterogeneity

Variability among demographic characteristics such as race, gender, and age within a population.

Real-world Applicability

The relevance of an algorithm's performance to actual scenarios and diverse populations.

Social Norms Evolution

Changes in societal standards and expectations that might affect the performance of algorithms over time.

Variations in Architecture

Different designs and structures of algorithms that can lead to discrepancies in performance.

Machine Learning Bias

Systematic errors introduced into machine learning models due to biased training data.

Equitable Performance

Ensuring that algorithms deliver fair and just results across diverse groups.

Scenario Analysis

Evaluating potential outcomes based on different conditions and variable changes in algorithm performance.

Racial Composition of Samples

The makeup of racial groups within datasets influencing algorithm performance.

Algorithm Implementation

The method by which an algorithm is executed within a system to achieve desired results.

Capture Diversity

Including various demographics in data collection to ensure representative training of algorithms.

Identification Discrepancies

Differences in how well an algorithm can identify individuals across different demographic categories.

Continuous Evaluation Process

The ongoing assessment of algorithms to ensure fairness and accuracy over time.

Quality of Training Photographs

The clarity and conditions of images used to train algorithms, impacting their future performance.

Robustness to Demographic Shifts

The ability of an algorithm to perform consistently despite changes in the demographic composition of the population.