info@embecosm.com

My colleague William Jones is Head of AI/ML at Embecosm, and often posts here about his work in Bayesian Inference machine learning. However he has another role, which is leading the TechWorks initiative to develop best practice guidelines when developing AI in electronic systems. This came to fruition recently, when William presented the first release of the guidelines at the launch of TechWorks AI at Bletchley Park.

The guidelines are based on the same principles as those developed by the IoT Security Foundation. They take the form of a set of questions, in ever finer detail to be addressed before and while developing an AI solution. As with the IoTSF guidelines, the document is freely available for all to use. The target audience is the experienced engineer, used to leading major projects, but with little prior experience of AI or machine learning.

At the top level, best practice guidelines require you to answer five questions about your project.

1. Should I use AI/ML?
2. If so what kind of AI/ML?
3. Where do I get my data?
4. How do I train my model?
5. How do I deploy my model?

It is surprising how many times, that first question is not asked. In our experience, many of the problems we look at do not need AI, just conventional automation.

The second question is often misunderstood. It is not just what sort of neural network I should you use, but are neural networks of any sort the right approach. If you need explainable answers, or need to find the likelihood of more than one possible answer, maybe you should be using a Bayesian technique such as Dynamic Causal Modeling. Or if you are trying to create code, perhaps Genetic Programming or Inductive Logic Programming would serve you better.

Most engineers understand the need for data to drive and machine learning based solution. But the importance of having the right data, in the right form is absolutely critical. It is the difference between an AI solution that works, and one that shows bias or hallucinates.

And actually the last two questions are the ones that everyone understands. But if you don’t get the first three right, then your project is doomed to failure.

The TechWorks Best Practice Guide for AI in Electronic Systems is a living document. It will evolve with the technology. You can see it here and contribute to its development, through pull-requests on its GitHub repository or by joining the monthly review call. I look forward to your contributions.