Wubin Bai | University of North Carolina: How do you scale neural networks for optical wearables when human anatomy and skin color vary so drastically?
00:15:31 - 00:16:51
Other snippets from this talk
Summary of the clip:
How do you scale neural networks for optical wearables when human anatomy and skin color vary so drastically?
When moving from one individual to another, differences in skin color, tissue density, and underlying muscle anatomy cause major variations in optical signal baseline readings. This variability makes static, population-wide machine learning models highly inaccurate.
To address this challenge, researchers implement transfer learning algorithms. Instead of retraining a massive neural network from scratch for every new user, a pre-trained model is used as a foundation.
By introducing a very small dataset of the new user's muscle movements, the transfer algorithm recalibrates and normalizes the network. This approach significantly reduces the barriers to scaling wearable diagnostics across diverse patient populations.
In this short video, you can learn:
* How biological and anatomical variability affects optical muscle-sensing accuracy across populations.
* The role of transfer learning algorithms in adapting pre-trained neural networks with minimal datasets.
* How normalization techniques make individual-focused AI models widely scalable.
π **Clip Abstract**
This Q&A segment addresses how to overcome the anatomical and optical variability of sensor readings across different users. By implementing transfer learning, pre-trained individual models can be easily generalized to new populations with very small training datasets.
π Link in comments π
#TransferLearning, #OpticalMyography, #SignalNormalization, #OpticalBiosensors, #WearableDiagnostics, #FlexibleElectronics
This is a highlight of the presentation:
Multi-modal noninvasive in vivo biosensing system to leverage sensor-algorithm synergy for enhanced accuracy
Future of Electronics RESHAPED USA 2026
10-11 June 2026
Computer History Museum, Mountain View, California, USA
Organised By:
TechBlick
More Highlights from the same talk.
00:02:18 - 00:03:13
How can light penetrate deep enough into tissue to map internal muscle deformation?
How can light penetrate deep enough into tissue to map internal muscle deformation?
Using optical wavelengths that target the biological window is essential for deep tissue penetration. Near-infrared (NIR) light is particularly effective because it can reach depths of several millimeters to a centimeter depending on the source intensity.
The mechanism relies on the muscle's high concentration of myoglobin, a protein structurally similar to hemoglobin in blood. Myoglobin strongly absorbs near-infrared light, meaning that when NIR light is projected into the muscle tissue, it undergoes significant absorption and modulation.
The backscattered light that escapes the tissue carries modulated information that corresponds directly to muscle structure. By placing flexible photodetectors on skin-mounted printed electronics, researchers can capture these dynamic reflections to track physical muscle contractions without invasive procedures.
In this short video, you can learn:
* Why near-infrared (NIR) light corresponds to a biological window allowing deep tissue penetration.
* How the muscle-specific protein myoglobin modulates NIR light absorption and backscattering.
* How structural muscle deformation can be captured by flexible photodetectors on printed electronics.
π **Clip Abstract**
This clip explains the optical physics behind non-contact muscle sensing using near-infrared light. By leveraging the specific absorption characteristics of myoglobin, the system tracks muscle deformation remotely through modulated backscattered light.
π Link in comments π
#NearInfraredSensing, #OpticalMyography, #FlexiblePhotodetectors, #EpidermalElectronics, #Biophotonics, #WearableBiosensors
00:05:31 - 00:06:35
Can optical wearable sensors overcome motion artifacts during intense physical activities?
Can optical wearable sensors overcome motion artifacts during intense physical activities?
Optical sensors are highly sensitive to local muscle movement, but overall body motion during activities like running or biking introduces severe motion artifacts. To isolate muscle-specific signals, a multi-sensor fusion approach is required.
By combining the optical sensor array with an Inertial Measurement Unit (IMU) containing a gyroscope, researchers can capture overall body orientation and movement. The global motion data acts as a reference channel to filter out noise from the optical data stream.
Through real-time data pre-processing, the motion artifacts are subtracted, leaving only the clean optoelectronic signatures. This allows the system to identify distinct, localized muscle signatures such as swallowing, deep breathing, and coughing even under intense physical stress.
In this short video, you can learn:
* How gyroscopes are used in wearable devices to track global body motion and orientation.
* The pre-processing steps required to cut off motion artifacts from optical sensor streams.
* How the multi-channel sensor array identifies unique signatures for specific throat and neck movements.
π **Clip Abstract**
This clip focuses on integrating IMU gyroscopes with optical arrays to isolate localized muscle movements from larger body motion artifacts. The resulting pre-processed signals yield highly distinct, clean biometric signatures for activities like swallowing and coughing.
π Link in comments π
#SensorFusion, #MotionArtifactReduction, #InertialMeasurementUnits, #OptoelectronicSensing, #WearableBiometrics, #FlexibleElectronics