E176 ‘Ethical AI and Bias Mitigation’ with University of Limerick’s Deepak John Reji

Today’s guest is Deepak John Reji, Senior Data Scientist, and Associated with University of Limerick. Discussing Deepak’s AI Awards 2024 nomination in the Best Use of Responsible AI an Ethics category, this episode centres around the challenges and opportunities in creating safe and responsible large language models (LLMs) that strive to reduce bias and bolster ethical AI practices.

Deepak introduces the topic by underlining the enormous challenge posed by bias in LLMs, which are often inherent in the datasets they are trained on. He explains the project’s genesis, which began in 2021, as an effort to detect bias-causing words and phrases and explore ways to mitigate these biases without degrading the model’s existing knowledge. This ambitious endeavor aims to ensure that AI solutions remain ethical and transparent, achieving better bias reduction and knowledge retention through a rigorous and diverse testing process.

Deepak highlights that creating unbiased AI models is a complex task. He explains that curating datasets involves collecting data from potentially biased internet sources. The subjective nature of bias perception based on geographical and cultural contexts adds another layer of complexity. To combat this, a scoring methodology was introduced, utilising feedback from a diverse group of annotations experts. This innovative measure ensures that varied social identities and values are integrated into the model’s refinement process.

Deepak also notes the technical hurdles of training large-scale models, which require immense computational resources and generate significant carbon emissions. The scrutiny of evaluation methods is equally crucial, given that models often behave differently across demographics and cultural contexts.

Upon open-sourcing their project, feedback from diverse audiences provided fascinating insights into the subjective nature of bias. Deepak shares an example which illustrates how a statement perceived as biased in one region may not evoke the same reaction elsewhere. This exercise underscores the importance of ensuring that datasets and their representations are universally accessible and free from unintended cultural biases.

Deepak believes that their approach to instruction-led fine-tuning can be adapted across various sectors, including healthcare, retail and education. While many companies rely on API models without internal customization, this method of tailoring solutions can provide clients with safer and more relevant AI applications. Importantly, he emphasises the significance of responsible AI usage from inception, advocating for ethical considerations to permeate the entire development process.

Deepak asserts that the commitment to ongoing improvement is key as the tech landscape is rapidly evolving. There is a constant need to adapt research and development to align with the latest technological advancements. He advocates for integrating cognitive science into AI systems to better understand human behavior and thought processes while engaging with these technologies. By doing so, developers can address the ethical dimensions that go beyond mere technical implementation.

In conclusion, this episode sheds light on the intricate challenges and promising opportunities that accompany the quest for responsible AI. Deepak’s thoughtful engagement highlights the collaborative spirit and multidisciplinary efforts necessary to advance ethical AI solutions. His insights serve as a crucial reminder that responsible AI development must be a foundational element of innovation, guiding the way for more inclusive and trustworthy AI applications worldwide.