Algorithms and Bias
Understanding Bias
- Definition of Bias
- Unfair beliefs or behaviors directed towards an individual or group.
- Typically tied to preconceived notions regarding gender, race, age, or sexual orientation.
- Can be positive (e.g., taller athletes being favored) or negative (e.g., racism).
- The historical context of bias in contexts like sports (e.g., Michael Lewis' "Moneyball").
Implicit Bias
- Definition
- Often unconscious bias; individuals may not be aware of it.
- Stereotype Threat
- Example experiment:
- Math test presented as challenging → Men outperform women.
- Math test presented neutrally → Performance equal across genders.
- The subjectivity and controversy surrounding the phenomenon.
Our Shared Biases
- Cultural Recognition
- Everyone possesses some level of bias.
Search Engine Bias
- Examples of bias in search engines:
- Ads for executive positions are displayed less to women.
- Searching "doctors" predominantly shows men, despite 34% being women.
- Racially biased outputs:
- Example searches show disparities in representation of groups.
- Notable improvements since awareness was raised.
AI and the Justice System
- Applications
- Use of predictive algorithms in parole, sentencing recommendations, and policing.
- Can perpetuate existing biases through historical data.
Bias in Robotics
- Concerns
- Life-or-death decisions made by military robots and system errors, particularly with vision.
- Medical robots needing to make ethical value judgments (e.g., in triage situations).
Comparing ML and Symbolic AI
- Definitions
- Machine Learning: Utilizing datasets with statistical methods.
- Symbolic AI: Processing symbols that hold real-world meanings.
- Future Trends
- Movement towards ML while also anticipating a need for synthesis with symbolic reasoning.
Understanding ML Bias
- Black Box Issue
- Algorithms typically provide answers without clear reasoning behind them.
- Concept of "garbage in, garbage out" reflects on data quality being paramount to result quality.
Explainable AI
- Current Research
- Focus on creating ML models that can elucidate reasons behind decisions made.
- Developing rule-based expert systems as a potential methodology for transparency.
Types of Bias in ML Datasets
- Key Bias Types
- Reporting Bias: Extreme opinions can skew understanding (e.g., book reviews).
- Automation Bias: Users placing undue trust in algorithmic outcomes.
- Selection Bias: Non-representative data selection.
- Coverage Bias: Skewed data collection methods leading to poor representation.
- Non-response Bias: Variation in response likelihood based on characteristics.
- Sampling Bias: Non-random sampling methods (e.g., first 200 responses).
Group Attribution Bias
- Definitions
- In-group Bias: Favoring individuals similar to oneself (e.g., alumni hiring).
- Out-group Homogeneity Bias: Stereotyping outsiders.
Types of Bias in ML Datasets (cont.)
- Additional Bias Types
- Implicit Bias: Assumptions based on personal beliefs that may not be widely applicable.
- Confirmation Bias: Tendency to validate personal beliefs unconsciously.
- Experimenter’s Bias: Iterative modifications to uphold expected results.
FATE Model in AI
- Elements
- Fairness: Ensuring equity in AI applications.
- Accountability: Maintaining responsibility for decisions made by AI.
- Transparency: Demanding clarity in algorithms and processes.
- Ethics: Ethical considerations in AI deployment and functionalities.