Implementation of a User Attention Detection System Based on Convolutional Neural Networks in Android
Thesis type IMPLEMENTATION, RESEARCH
Description Ever since their introduction, smartphones have become more and more advanced and powerful. This however comes at the cost of the device’s battery life. Currently, power management in smartphones is mostly based on simple static policies (such as timeouts) that do not take usage context into account, except for very simple cases (such as turning off the screen when the device is in a pocket).
Potentially, much more advanced context-dependent power management policies could be implemented using the large number of sensors available in modern smartphones.
The goal of this thesis is to evaluate the feasibility of implementing one of such policies on Android. Specifically, the policy is based on a deep learning classifier, able to determine if a device is “in use” or not based on low-power sensors, and perform power management actions accordingly (such as turning of the screen or changing Doze state).
A first version of the classifier has already been developed and tested in Python (PyTorch) during a previous project, using data gathered from several smartphones. Consequently, the first part of the thesis will focus on implementation. Specifically, the student will port this classifier to an Android service executed on the actual device, and evaluate its overheads in terms of execution time and power consumption.
After that, the student will work on improving the classifier both in terms of improving its accuracy and reducing its complexity. This second part of the thesis will be more research related and may lead to a scientific publication.
Required skills Canditates should have good programming skills and should be interested in machine/deep learning applications.
Previous experience on Android and/or Python and basic notions of machine/deep learning are desired but not required.
Deadline 24/10/2020 PROPONI LA TUA CANDIDATURA