Determining the potential of your AI algorithm starts with measurement

This article originally appeared in the August'22 magazine issue of Electronic Specifier Design – see ES's Magazine Archives for more featured publications.

To what depth a system can measure can be thought of as its potential. Determining what aspects of those measurements must be sent to the processor can be thought of as delivering that potential. Finally, knowing how to combine the salient parts of those measurements in the correct proportions, known as sensor fusion, is the key to exploring an algorithm’s IQ or reasoning potential. Augment that sensor fusion with a feedback loop and the algorithm will have the ability to check and course-correct its logic, a necessary ingredient in machine learning. By Jeff Harris, Vice President, Global Corporate and Portfolio Marketing, Keysight Technologies.

These three attributes are the key to understanding the depth of an AI’s unique power. And like many things, the more you cultivate and calibrate these foundational elements, the better the AI algorithm will perform in the long term. Now that we understand the three areas to explore let's dive into the first component, measurement depth, and how it’s critical to the foundation of building a robust, high-performing AI algorithm.

Measurement depth

Metrology is the study of measurement science and measurement depth plays a crucial role in building a robust algorithm. The Gaugemaker’s Rule, or 10:1 rule, states that a measurement device must be 10x more precise than the desired measurement. The reason that measurement depth is so critical is that it determines the possible level of precision and sets the algorithm's maximum potential. Therefore, the more precision you have in any given measurement, the greater the AI algorithm’s potential. 

Metrology focuses on the deep understanding of a particular measurement. That measurement can be as simple and distinct as voltage, ground, or temperature or as multi-modal as the functioning of aircraft control surfaces, or as complex as maximising throughput on a manufacturing assembly line. Whether you are measuring a single parameter or several, the depth of each measurement determines the level of programmability that’s possible. For instance, measuring a 3 Volt system to 1/10th of a volt is not as insightful as measuring to 1/1000th of a volt.

Depending on the system that voltage is powering, the extra precision may be critical for battery life or maybe a distraction. Maximising the potential of any algorithm requires matching the entire end-to-end measurement needs to the needed depth. This is true no matter what’s being measured, even data systems, which may not be as immediately intuitive, so let’s look at one of those examples. 

How to optimise measurement

Enterprise IT stacks are now a complex web of interconnected data systems, each exchanging information aimed at tuning an organisation’s operations. These technology stacks include an array of software such as CRM, ERP, databases, order fulfilment, and each with unique data formats and custom application programming interfaces (APIs).

According to Salesforce, the average company has over 900 applications in its tech stack, many of them cloud-based and all of them experiencing software updates that can have ripple impacts. Identifying and isolating problems, much less optimising performance across multiple intersecting software applications, is akin to finding a needle in a collection of interconnected haystacks.

Each software application in a tech stack has a different sponsor in an enterprise – finance, human resources, sales, marketing, supply chain – and that primary org’s considerations are top of mind for IT. Every enterprise has custom workflows and integrations with numerous applications and backend systems, and user journeys span various paths and are rarely linear. Therefore, even if two enterprises used identical applications in their tech stack, mapping all the exchange points and validating the end-to-end operation would be unique. If there were ever an application in need of AI, this would be it. The measurements, in this case, could be the intersystem data input points, the intrasystem data exchange points, and the data display points.

Understanding how an AI algorithm would operate in a system like this would start with understanding how it measures points data in three key areas:

1. Measuring how users’ interface with the application, regardless of the operating system, which in some cases involves employing robotic process automation (RPA) when button pushes are required

2. Measuring the data exchanges between and command APIs that link the systems in a complex technology stack to ensure they are occurring correctly 

3. Measuring on-screen information across omni platforms (desktops and mobile) such as images, text and logos as a human would to see how they render

Evaluating the measurement efficacy starts with its ability to measure regardless of operating system, software versions, devices, or interface mechanisms. The more conditions under which the AI cannot measure, the less impactful it will be in operation.

Conclusion

Whenever you assess the potential of anything, start with the foundation. At the foundation of every AI system is its ability to measure. The more it can measure, the more impactful it has the potential of being. Look at what it is capable of measuring and, more importantly, where it is not capable. Limited sensing results in limited AI algorithm potential. The old adage from Lord Kelvin stands true today that “if you cannot measure it, you cannot improve it.” To understand the true power of any AI, make sure to start by analysing its measurement breadth and depth.