Improving the LYS Button with AI-driven Discrepancy Modeling

16 Nov 2022

Hi, I’m Forrest Webler, Ph.D. and Chief Science Officer at LYS. I’m excited to share our biggest update to the LYS Button in years. In this update, we’ve improved the accuracy of the LYS Button, introduced new values like mEDI, and updated our tech specs with useful resources for grant writing, and data analysis just to name a few.

Metrology (the science of measurement) has always fascinated me. In fact, one of my first published papers is written on this topic. The idea that humans associate symbols with physical quantities and phenomena are, on its own, truly remarkable. Furthermore, the ability to claim, beyond a reasonable doubt, that two things are indeed different could be argued to be the origin of science itself. Thus, the ability to measure quantities precisely is a prerequisite for discovery.

Early description of a cubit rod (one of the earliest measuring devices) from the Museo Egizio of Turin
Early description of a cubit rod (one of the earliest measuring devices) from the Museo Egizio of Turin

At LYS we seek to enable discovery in fields as diverse as sleep and circadian neuroscience to early childhood development and dementia care by providing a tool for measuring light. Officially launched in 2017, the LYS Button has been widely adopted by researchers and practitioners alike across 53 countries including some of the world’s most prestigious institutions including Harvard Medical School, University of Oxford’s Nuffield Department of Clinical Neurosciences, and the WELL living lab. In light of our ongoing commitment to all our clients, we strive to improve our products.

LYS Technologies

When I joined the LYS team as CSO, I wanted to revisit the Button a half-decade on and see if we could apply new methods in signal processing and artificial intelligence to improve the resolution and type of data output from the sensor. Measuring light may seem intuitive, yet limitations in electronic component manufacturing place strict limits on the amount of information that can be extracted from the types of compact integrated circuits we use to design small compact wearable sensors. I wrote a paper deep diving into this topic that can be accessed freely here. To circumvent these challenges, we have implemented changes to the software that processes the raw output from the Button.

The approach we took is based on Discrepancy Modeling. In many physical systems, there are sometimes observations that fall outside of a mathematical model much to the chagrin of engineers. While these discrepancies have been historically written off as errors, applied mathematicians have since questioned if there is value in attempting to find systematic patterns within the errors to improve model accuracy. While this may sound intuitive, the difficulty in isolating hidden deterministic patterns in seemingly random residuals was immense prior to the ubiquity of deep learning today.

In the case of the Button, we wondered if we could account for some of the observed variation in the illuminance measurements by adding a discrepancy model to our existing model. Since the Button does not measure illuminance directly we have to compute it from the tristimulus channel data.

Where Ev is illuminance in lux, f is a function of the raw sensor output x = [R, G, B] and E is the measured error. With discrepancy modeling, we can “absorb” some of the error into a new discrepancy term by identifying unknown additional relationships between the channel outputs and measure illuminance. In this case, our model becomes

where N is the remaining unexplainable error or signal noise. Because we now have a model g that accounts for a portion of the original error, we can improve our measurement accuracy leaving the only unknown variable of noise as a source of variation in our measurement accuracy.
While this approach may appear deceptively simple, the challenge rests with finding the discrepancy term. Unlike f which is typically provided by the sensor manufacturer, the discrepancy term g can be a complex non-linear relationship between inputs and desired target outputs. For this reason, researchers in the field of discrepancy modeling have promoted the use of deep neural networks. We go one step further to maximize interpretability by adding a step called Symbolic Regression that essentially “converts” our learned neural network model into an interpretable equation that can be easily coded using conventional scientific computing software.
This approach means that the more data we collect with LYS Buttons the better and more accurate they will become. By updating our discrepancy model periodically we can ensure that we are living up to our mission of delivering the highest quality and innovative product solutions to all our customers.

All relevant technical documentation can be found in our Technical Specifications – download them here. If you want to try out the LYS Button, please find our research sample package here.

I personally hope you will enjoy our latest update and please contact us if you have any questions or book a demo of the LYS Button.

About the LYS Button

Since 2017, the LYS Button has been a reliable and compact tool for researchers and practitioners alike. As our first product, the Button remains near and dear to our hearts at LYS. It reminds us of the importance of sharing awareness of the light environment to educate about the need for healthier light indoors and daily exposure to natural light outdoors.