Systems and methods for responsible AI
Systems and methods for responsible artificial intelligence (AI) may provide an end-to-end framework from defining what responsible behavior and guidance are to the components that developers can leverage to detect and measure bias. API components may be provided at the model training and deployment stage to mitigate bias. Deployed components may then be linked to fairness monitoring APIs in a machine learning operations framework to integrated in the enterprise.
1 . An end-to-end method for responsible artificial intelligence (AI) using a platform implemented on one or more hardware computer processors and one or more storage devices, the method comprising:
in a data privacy step, inputting data in an input database capable of receiving data;
in a data bias step, using a data bias system configured to execute code, conducting exploratory data analysis (EDA), pre-processing to identify missing or incomplete records, outliers and anomalies, and/or perform feature engineering and selection on the data;
in an explainable artificial intelligence (XAI) and privacy step, using a model development system configured to execute code, developing a model through XAI and privacy;
in a model bias and privacy step, using the model development system, evaluating and selecting the model for bias and privacy;
evaluating prediction bias of the model;
deploying the model; and
following deployment, using a unified model monitoring platform (UMMP) comprising a model deployment layer; a gateway layer that provides access to the data through an application programming interface (API) or an event stream; a risk analysis index (RAI) services layer including a plurality of independent microservices; a RAI APIs layer that provides access to the RAI services layer to extract RAI details for the model, generate a model benchmark report, subscribe to violation notifications, and get the model in production; and a RAI user interface layer that provides a browser-based interface and includes at least an RAI dashboard, report generation, monitoring, and administration, monitoring the model to evaluate bias, XAI, and drifts.
2 . The method of claim 1 further comprising:
in a model accountability step, managing the model before deployment.
3 . The method of claim 1 , wherein following the monitoring step, selectively returning to the defining or data bias step to further refine the model.
4 . The method of claim 1 , the data bias step further comprising
selecting one or more sensitive features;
conducting bias measurement; and
if bias is detected, performing data debiasing.
5 . The method of claim 4 , the evaluating prediction bias of the model step further comprising:
selecting one or more sensitive features;
conducting model debiasing; and
measuring bias.
6 . The method of claim 5 , wherein the model debiasing step occurs following in-processing or post-processing.
7 . The method of claim 4 , the monitoring step further comprising:
evaluating if bias is detected; and
if no bias is detected in the evaluating step, confirming model privacy, resulting in a fair and secure model.
8 . The method of claim 7 , wherein model privacy is confirmed through one or more of the following:
bias and fairness drift, model monitoring, data drift, and/or model drift.