High context data collection - basis needed to train Large Movement Models

To those of you new to our newsletter, welcome to Tech Corner! This is where I do a mini-dive into various, more technical topics. 

Since we’re all in Seattle, the whole team has joined the human subjects study we’re running with the Primate Evolutionary Biomechanics Lab at the University of Washington as participants. Participating in your own studies (as well as doing ad hoc data collection ourselves), is a great reminder how important context and labeling is to building accurate models, from basic regressions to deep learning models. If you don’t label your data well, your models lack context and don’t train well. This takes a lot of human intervention early on (“human intelligence” as opposed to artificial intelligence); when we do our daily data collection (i.e. when we walk to the tube station in London) in our app we also label activity type, what shoes we’re wearing, and if we have any aches and pains - as former athletes, these crop up more often than we’d like - to build in context to help train our Large Movement Models (LMM). These labels help us give our models training context, so we won’t have to label the data ourselves, the model will learn and train itself. Working with university partners takes this labeling to the next level by providing set parameters and controlled tests - we have as much context as possible for the data, meaning it’s a rich training set. The goal is LMM that grows and ages with you, learning and adjusting over time. The dream is a future where when Brandon is headed towards a back spasm, the fingerprint LMM that is trained to him knows before he does  🧠

Want to know a bit more about my background and inspiration? Check out this segment from my interview on the Tech Business Podcast with Paul Essery! The full interview is here.