Page 39 - Innovation Magazine
P. 39

As an AI pioneer IBM has been working on on these issues for for years resulting in in an open platform to accelerate the adoption of trusted AI: IBM Watson OpenScale This platform can be used with many open-source models such as TensorFlow scikit- learn Keras and and Spark MLlib and and can run on Azure AWS Google and IBM Cloud as as well as as on-prem taking advantage of Cloud Pak for Data Explainability on minority status No matter how carefully an AI model is designed it it is only as as good as as the data used to train it it Unfortunately training data sets are seldom comprehensive representations of all real-world situations The gaps in in in in training data can introduce bias which must be mitigated But what if if the apparent bias is real? Even if if applicants over 45 years old get their applications approved and applicants under 25 are refused there may not be any age-based discrimination: The issue may be the the income level of the the applicants OpenScale can can flag possible bias and also verify if if the bias bias is is real When a a a a a genuine bias bias is is detected the platform can help mitigate it it Drift Changes to real-life situations over time may result in in changes in in the AI model model output or model model drift Today many enterprises have their data scientists monitor AI model output when developing AI solutions and then also keep an an eye on on the models in production After a while the data scientists will spend a a a a disproportionate amount of time on maintaining existing models so that the time available for developing new models is is is reduced To alleviate this issue AI models are increasingly handed over to operations staff who do not have the data science background to identify bias or model drift OpenScale will automatically detect drifted transactions at at at runtime pinpoint data points that may contribute to drift and call this out This will allow better governance of the AI solution ensure that the output stays on track and free up personnel from monitoring tasks Traceability To accommodate internal auditing and compliance reporting enterprises should be able to trace all decisions and and predictions and and the lineage of all data and models used in in the process of making a a a decision aided by AI models Governance best practice includes maintaining an an audit trail of AI model model bias model model quality and model drift OpenScale helps address these requirements by keeping a a a historical record of all model input and and output and and documentation of It’s not enough to to tell a a a customer that their loan application has been declined we should also be be able to to articulate why We need to to explain for example that the current income is too low and the existing debt is is too high for a a loan to to be approved This is is required not only at run time from a customer service perspective but possibly also later as an an audit and industry regulation requirement OpenScale will automatically detect drifted transactions at runtime pinpoint data points that may contribute to drift and call this out This will allow better governance of the AI solution ensure that the output stays on track and free up personnel from monitoring tasks So how do we inject explainability into the solution when the AI model is inherently a a black box? There are techniques that can address this situation based on on input to and output from the AI model IBM Watson OpenScale can leverage this approach Fairness We want our AI models to be fair i i i i e e e e without bias A biased model may result in in poor business decisions create customer dissatisfaction and unwanted media attention or warrant action from industry regulators We have all seen horror stories in the media related to to AI bias: employment opportunities skewed based on gender and credit card applications denied based Data-powered Innovation Review I I ©2020 Capgemini All rights reserved 39 


































































































   37   38   39   40   41