New Responsible ML innovation in Azure Machine Learning
Published date: 19 May, 2020
As organisations look to adopt artificial intelligence (AI), they face significant challenges in developing and using AI responsibly. To help organisations overcome this barrier, we are bringing the latest research in responsible AI to Azure, in collaboration with the Aether Committee and its working groups. The new responsible ML capabilities in Azure Machine Learning and our open-source toolkits empower data scientists and developers to understand machine learning models, protect people and their data, and control the end-to-end machine learning process.
- Understand: Model interpretability capabilities in Azure Machine Learning and fairness assessment and mitigation capabilities using Fairlearn enable the development of more accurate and fairer models.
- Protect: The new differential privacy WhiteNoise toolkit can be used with Azure Machine Learning to enable customers to build machine learning models using sensitive data, while safeguarding the privacy of individuals. This is a result of the partnership between Microsoft and researchers at Harvard’s IQSS and School of Engineering. Confidential machine learning capabilities enable data science teams at Microsoft to build models over confidential data in a secure environment, without being able to see the data. We will bring these confidential machine learning capabilities to developers and data scientists later this year.
- Control: Azure Machine Learning provides capabilities to automatically track lineage and maintain an audit trail of ML assets to meet regulatory requirements. Datasheets provide a standardised way to document ML assets, and provide transparency to data scientists, auditors and decision makers. Developers and data scientists can use custom tags in Azure Machine Learning to implement datasheets for models today.
These new Azure Machine Learning and open-source toolkit innovations have been built on decades of research and provide organisations with a comprehensive set of capabilities to develop AI solutions responsibly.