microsoft/responsible-ai-toolboxInteractive dashboards and libraries for responsible AI model debugging, covering error analysis, fairness, interpretability, counterfactuals, causal analysis, and data balance.
Responsible AI is an approach to assessing, developing, and deploying AI systems in a safe, trustworthy, and ethical manner, and take responsible decisions and actions.
Responsible AI Toolbox is a suite of tools providing a collection of model and data exploration and assessment user interfaces and libraries that enable a better understanding of AI systems. These interfaces and libraries empower developers and stakeholders of AI systems to develop and monitor AI more responsibly, and take better data-driven actions.
The Toolbox consists of three repositories:
| Repository | Tools Covered |
|---|---|
| Responsible-AI-Toolbox Repository (Here) | This repository contains four visualization widgets for model assessment and decision making: 1. Responsible AI dashboard, a single pane of glass bringing together several mature Responsible AI tools from the toolbox for a holistic responsible assessment and debugging of models and making informed business decisions. With this dashboard, you can identify model errors, diagnose why those errors are happening, and mitigate them. Moreover, the causal decision-making capabilities provide actionable insights to your stakeholders and customers. 2. Error Analysis dashboard, for identifying model errors and discovering cohorts of data for which the model underperforms. 3. Interpretability dashboard, for understanding model predictions. This dashboard is powered by InterpretML. 4. Fairness dashboard, for understanding model’s fairness issues using various group-fairness metrics across sensitive features and cohorts. This dashboard is powered by Fairlearn. |
| Responsible-AI-Toolbox-Mitigations Repository | The Responsible AI Mitigations Library helps AI practitioners explore different measurements and mitigation steps that may be most appropriate when the model underperforms for a given data cohort. The library currently has two modules: 1. DataProcessing, which offers mitigation techniques for improving model performance for specific cohorts. 2. DataBalanceAnalysis, which provides metrics for diagnosing errors that originate from data imbalance either on class labels or feature values. 3. Cohort: provides classes for handling and managing cohorts, which allows the creation of custom pipelines for each cohort in an easy and intuitive interface. The module also provides techniques for learning different decoupled estimators (models) for different cohorts and combining them in a way that optimizes different definitions of group fairness. |
| Responsible-AI-Tracker Repository | Responsible AI Toolbox Tracker is a JupyterLab extension for managing, tracking, and comparing results of machine learning experiments for model improvement. Using this extension, users can view models, code, and visualization artifacts within the same framework enabling therefore fast model iteration and evaluation processes. Main functionalities include: 1. Managing and linking model improvement artifacts 2. Disaggregated model evaluation and comparisons 3. Integration with the Responsible AI Mitigations library 4. Integration with mlflow |
| Responsible-AI-Toolbox-GenBit Repository | The Responsible AI Gender Bias (GenBit) Library helps AI practitioners measure gender bias in Natural Language Processing (NLP) datasets. The main goal of GenBit is to analyze your text corpora and compute metrics that give insights into the gender bias present in a corpus. |
Responsible AI dashboard is a single pane of glass, enabling you to easily flow through different stages of model debugging and decision-making. This customizable experience can be taken in a multitude of directions, from analyzing the model or data holistically, to conducting a deep dive or comparison on cohorts of interest, to explaining and perturbing model predictions for individual instances, and to informing users on business decisions and actions.
In order to achieve these capabilities, the dashboard integrates together ideas and technologies from several open-source toolkits in the areas of
Error Analysis powered by Error Analysis, which identifies cohorts of data with higher error rate than the overall benchmark. These discrepancies might occur when the system or model underperforms for specific demographic groups or infrequently observed input conditions in the training data.
Fairness Assessment powered by Fairlearn, which identifies which groups of people may be disproportionately negatively impacted by an AI system and in what ways.
Model Interpretability powered by InterpretML, which explains blackbox models, helping users understand their model's global behavior, or the reasons behind individual predictions.
Counterfactual Analysis powered by DiCE, which shows feature-perturbed versions of the same datapoint who would have received a different prediction outcome, e.g., Taylor's loan has been rejected by the model. But they would have received the loan if their income was higher by $10,000.
