Last year, Meta AI released the Casual Conversations dataset – an open-sourced data set consisting of 45,186 videos of participants having nonscripted conversations. The purpose was to help AI researchers find meaningful signals that can help them assess the fairness of their computer vision and audio models across gender, age, apparent skin tone, and ambient lighting subgroups.
Taking a step forward, now the team has released human transcripts from the Casual Conversations Dataset to encourage the community to come up with a variety of strategies for reducing statistical biases such as word error rate (WER) in Automated Speech Recognition (ASR) systems.
Till now, only a few research have looked at how well speech recognition models function for diverse individuals. The addition of human speech transcriptions will open the gate for researchers to utilise the shared dataset for measuring the performance gaps of the speech recognition systems across different groups of people.
“AI systems need data and benchmarks in order to measure performance, and the research community simply hasn’t had adequate ways to assess fairness concerns for speech recognition systems,” as mentioned in the blog.
Fairness beyond speech technology
It’s a well-known fact it was Joy Buolamwini’s research that helped persuade big tech giants IBM, Amazon and Microsoft to put a hold on their facial recognition technology. She is currently the founder of The Algorithmic Justice League (AJL) – an organisation working towards mitigating AI bias and harm. Earlier, biases against women in Apple’s credit card algorithm came into the limelight when Apple’s co-founder Steve Wozniak was given a credit limit 10 times higher than his wife, despite both sharing all assets and accounts.
It would not be an exaggeration to note that the fairness of AI algorithms is a burgeoning subject of study that stems from the requirement for choices to be free of bias and discrimination. Fairness also applies to AI-based decision-making tools, with the European White Paper on AI providing a framework for AI or algorithmic decision-making to be carefully considered.
Let us use a hypothetical case for the sake of simplicity: an AI model deployed at a bank is used to predict the loan eligibility of an individual and the prediction is based on the risk of default. Some of the notable points in the White Paper on AI by the European Union consider: a person should not be subjected to automated decision making in the first place, it is their right to have an explanation as to how the model comes to a conclusion, and lastly non-discrimination.
To be precise, this simply calls for the production of AI models that are fair (unbias), interpretable (explainable to end-users) and transparent – by design. Leading by example, several tech players have in-house research teams, products, and tools in this direction. Take, for example:
- Meta AI uses Fairness Flow, a diagnostic tool that enables their teams to analyse how some types of AI models and labels perform across different groups.
- IBM’s AI Fairness 360 (AIF360) is a comprehensive open-source toolkit of metrics for detecting and mitigating undesired bias in datasets and machine learning models, as well as cutting-edge techniques to counteract such bias.
- Google’s What-If Tool explores a models’ performance on a dataset, is working on ‘Responsible AI with Google Cloud,’ ‘Responsible AI with TensorFlow,’ and laid out general best practices for AI.
“Algorithms will continue to reflect and reinforce preconceptions that hold society and business back unless a determined effort is made”
Unreliable findings from biassed AI models have the potential to harm reputation, severely impact end-users and will lead to a scenario where people’s trust in AI models will be lost.
Although the bulk of biases emerges during the training of AI models, many unintentional biases emerge over time, necessitating developers to monitor their AI systems in real-time. It is also vital to test models in a real-world setting in order for them to perform better in the environments in which they are designed to perform.
Source: indiaai.gov.in
