Academic Innovations

India is at the cusp of an AI revolution with a significant portion of the artificial intelligence deployment being in the domain of user-interfaced systems. These platforms allow humans to interact with different AI and internet based computing services through touch, vision or sound based interfaces and are built using very advanced machine and deep learning algorithms. These softwares are deployed on a large scale and have the potential to impact over a billion individuals throughout the country, thanks to low costs and easy to deploy plug-and-play design. As these systems are deployed for use at such a huge scale, any biases or discriminatory performance can snowball very quickly to impact a large number of people. Most of the systems deployed in society are of a black-box nature and the industrial corporations developing these models do not share any information regarding the training data or model architecture or parameters. This makes it impossible for academic or third party researchers to identify and mitigate the noted biases. Thus, audits of such AI systems allow internal or third party researchers to identify the presence of existing and newly emerging biases. These biases can manifest in the following ways-- (i) misclassification or incorrect face matching by a Face Recognition System (FRS), (ii) incorrect transcript generation by an Automatic Speech Recognition System (ASR), etc. Existence of biases against marginalized or minority communities, specially from historically discriminated groups in the Global South countries has far reaching ramifications like denial of service, opportunities and facilities. These biases can be addressed using pre-/in-/post-processing strategies. Thus, any user-interfaced system must be continually audited to better understand and mitigate these biases. For this purpose, researchers use benchmark datasets composed of demographically diverse and well annotated data points that ensure a thorough analysis across multiple dimensions of age, gender, race, ethnicity, caste, geography, etc. In this project, we propose a three part pipeline for the audit, analysis and mitigation of biases in Face Recognition Systems and Automatic Speech Recognition Systems that correspond to a standard responsible AI pipeline.