1. ML Approach to Approximating Ambient Light Exposure
Lori Liu, Chau Vu, Brent VanZant, Jaikrish Chitrarasu, Michael Ostertag, Tajana Rosing
Rongrong Liu, Chau Vu, Brent Vanzant, Jaikrish Chitrarasu, Michael Ostertag, and Tajana Rosing  
To change this poster, replace our sample content with your own. Or, if you’d rather start from a clean slate, use the New Slide button on the Home tab to insert a new page, then enter your text and content in the empty placeholders. If you need more placeholders for titles, subtitles or body text, copy any of the existing placeholders, then drag the new one into place.
Methodologies
Introduction
Methodologies
Results
Background: 
An accurate measure of light exposure is crucial for circadian rhythm research
Participants must wear cumbersome devices during experimentation
Objective: 
Eliminate the need for person-bound sensors
Approach: 
Machine learning model trained with external inputs and features
Goal: Develop machine Learning model that best predicts the wearable sensor readings based on inputs from weather data, stationary sensor, GPS, and acceleration.
Data consists one set of data where the wearable sensor was inside, and another with the wearable sensor outside. A mixed data set is used to test and evaluate the accuracy of our model.
Materials and Features
AS7262 Light Sensor 
SD Card Module
Data Collection
Stationary Sensor placed in frequently dwelled living space
Participant wears portable sensor during day-to-day activities
GPS and Acceleration recorded on participant's phone
Feature Extraction
Raw GPS coordinates:
 - Indoor vs Outdoor Indicator
 - Proximity from Stationary Sensor Indicator
Acceleration values:
 - Motion vs Sedentary (contextualize indoor vs outdoor)
- (top right) Explained variance score over the inside/outside/mixed dataset.
- (bottom) One example of predicted value compared with prediction value for the inside dataset.
9V Battery
AS7262 Light Sensor 
Conclusion and Furture Work
SD Card Module
Arduino
Micro
We assembled one wearable sensor and one stationary sensor for data collection
- We experimented with several model structures to find one to best fit. We will continue to investigate better combinations of features.
- In the future, we will expand the variety of the wearable dataset to include routes outside UCSD campus and validate on the model's accuracy.
(autonomous device/ algorithm/diagram)
Feature
Source
Data
Outdoor light ambience
SDSC Weather Data
Dark Sky API
Real-time weather data with UV, solar radiation, wind, etc.
Indoor light ambience
Stationary Sensor
6-channel light spectrums (lux)
Location
GPS Phone Application
Latitude and longitude coordinates
Acceleration
Google Science Journal App
Acceleration in x/y/z axis
Ground truth light exposure value
Wearable Sensor (pictured above)
input
Arduino
Micro
9V Battery
Weather
Stationary
Perceptron
Prediction
MLP
Random Forest
Indoor vs outdoor
Weather
Stationary
Perceptron
Prediction
MLP
Random Forest
Indoor vs outdoor
MLP
Acknowledgement
This project is hosted by Early Research Scholar Program of Computer Science Engineering Department. We would like to express our deep gratitude to Professor. Alvarado, the program director, and Professor. Rosing, our research advisor, for patient guidance. We thank Michael Ostertag, our project supervisor, for guiding project development on a weekly basis, and Vignesh, the program assistant, for keeping our progress on schedule. 
Training the Model
Mixture model to predict light exposure :
Multi-Layer Perceptron model for Weather Data
Random Forest model for Stationary Data
Indoor vs Outdoor indicator for weight distribution
output
Multi-Layer Perceptron