Expert Insights

Alan Buxton emphasizes the criticality of handling 'out of distribution data' in AI-assisted software development, a challenge often overlooked with AI implementations. He highlights the complexities associated with encountering data that the AI model has not been trained on, indicating that robust testing mechanisms are necessary to manage such instances.

Alan suggests that leveraging traditional software guardrails can help address these challenges, underscoring the need for established software practices in the emerging field of AI.

Here's what Alan shares:

  • The problems AI models face when dealing with out-of-distribution data.
  • The importance of traditional testing methods in identifying and rectifying these issues.
  • Examples where traditional software practices have helped in dealing with the challenges posed by unseen data in AI models.

Quote

quotation-marks icon
So a, a couple of examples where we're doing that. So in our spend classification piece, for example, um, uh, language models, one of the big, one of the problems with a lot of this AI stuff is how it handles what we call out of distribution data. So data that it's never been trained on before.quotation-marks icon
Alan Buxton ,
CTO, Simphony

THE NEW DEFAULT angle

Here are a few actionable steps to face the challenge of handling 'out of distribution data' leveraging traditional software guardrails:

  • Build a solid preparatory framework. Ensure your AI model is trained with a wide range of data to minimize encounters with unseen data.
  • Integrate existing testing methodologies. Keeping traditional software testing methodologies intact alongside new AI models can help anticipate and manage new data effectively.
  • Encourage continuous learning within the team. Equip your teams with the knowledge and skills to handle the unpredictability of out-of-distribution data.
  • Iterate and improve. Regularly review your AI model's performance with new data to enhance its adaptability and predictability.
  • Be proactive in addressing anomalies. When novel data is encountered, use these instances as opportunities for model improvement, rather than setbacks.