Tiếp theoHướng dẫn tiếp theo
Cách Thực hiện Kiểm toán Thiên vị AI
kỹ thuật
HƯỚNG DẪN KỸ THUẬT
Bias mitigation can intervene before training, during model learning, or after predictions.
Nghe hướng dẫn nàyAI thuật lại · ~3 min
Pre-processing changes data or weights; in-processing changes the learning objective or algorithm; post-processing adjusts outputs or thresholds. Each method targets particular metrics and assumptions, so reducing one disparity can trade off against accuracy, calibration, or other forms of fairness.
Mitigation methods are often grouped by where they intervene in the machine-learning pipeline. Pre-processing changes the training data or sample weights before a model is fit. Reweighing, for example, changes how instances contribute to learning without necessarily altering the original rows. In-processing changes the learning procedure by adding a fairness constraint, penalty, or adversarial objective. Post-processing changes predicted scores or decisions after a base model is trained, such as choosing thresholds to reduce a selected group disparity. These categories describe mechanics, not guarantees. A method can optimize one fairness definition while worsening another. Demographic parity, equalized odds, and calibration answer different questions and may conflict when base rates differ. The appropriate goal depends on the decision, data, law, and affected people. A threshold adjustment can create different treatment across groups; it may also raise legal or operational questions. Reweighing cannot fix a target that measures an unjust outcome, and adversarially removing group information may not remove correlated proxies. Toolkits implement specific methods. AIF360 provides metrics and pre-, in-, and post-processing algorithms, including Reweighing. Fairlearn includes disaggregated assessment and mitigation techniques such as ExponentiatedGradient and ThresholdOptimizer. ThresholdOptimizer can apply group-specific thresholds under a selected constraint and objective. The method assumes access to sensitive features for fitting or prediction and may use randomized decisions in some configurations. Tools do not choose the legally or socially appropriate fairness criterion for a team. A sound workflow first defines the harm and metric, then establishes a baseline, applies a method, and evaluates on held-out data across relevant groups and intersections. Report accuracy, calibration, uncertainty, and operational consequences alongside fairness metrics. Document which tradeoffs were chosen and who approved them. If no acceptable solution meets safety, validity, and legal needs, do not treat a toolkit result as permission to deploy.
A team uses AIF360 Reweighing to assign different training weights to records so groups contribute differently without editing each feature value.
A scikit-learn workflow uses Fairlearn ExponentiatedGradient with a chosen constraint and checks the resulting accuracy and subgroup metrics.
A hospital applies a post-processing threshold method to model scores, then checks whether group-specific thresholds are clinically and legally appropriate.
A text classifier adds an adversarial objective to reduce how much a representation reveals about dialect, then tests whether task performance and other error patterns change.
These projects update APIs, supported algorithms, and defaults. Pin tested versions and reproduce metrics after upgrades; do not assume a new release preserves a result. The fairness objective remains a social and legal decision rather than a software default. Reassess when populations, uses, or rules change, and have domain owners approve any new constraint or group-specific decision rule before deployment. Treat tool output as evidence, document unresolved tradeoffs, and limit or stop use when required safety or validity standards are not met.
Bản tin miễn phí
Ba câu chuyện AI đã được xác minh vào mỗi buổi sáng các ngày trong tuần, được viết bằng tiếng Anh đơn giản. Miễn phí mãi mãi, không có quảng cáo.
Một email mỗi ngày trong tuần. Hủy đăng ký trong một cú nhấp chuột. Chúng tôi không bao giờ bán hoặc chia sẻ địa chỉ của bạn.
Tự kiểm tra
Phản hồi tức thì về mọi câu trả lời và chứng chỉ có thể chia sẻ với ID có thể xác minh sau khi bạn vượt qua khóa học.
Hỗ trợ giáo dục AI miễn phí. AI Understanding là tổ chức phi lợi nhuận 501(c)(3) — không bao giờ có quảng cáo, không có tường phí. Hãy quyên góp
Bias mitigation can intervene before training, during model learning, or after predictions. Pre-processing changes data or weights; in-processing changes the learning objective or algorithm; post-processing adjusts outputs or thresholds. Each method targets particular metrics and assumptions, so reducing one disparity can trade off against accuracy, calibration, or other forms of fairness.
Pre-processing changes data or weights before model training.
Post-processing changes model outputs or thresholds after a base model is trained.
Reweighing changes instance weights so they contribute differently during training.
ThresholdOptimizer uses sensitive features to fit or apply group-specific thresholds under specified constraints.
Different fairness criteria address different goals and may conflict.
Tiếp tục học hỏi
Đã chọn thêm hướng dẫn cho chủ đề này
Tiếp theoHướng dẫn tiếp theo
Cách Thực hiện Kiểm toán Thiên vị AI
kỹ thuật