Designing Compressed Deep Neural Networks for Inappropriate Scene Detection
7th October 2019
Proposed by Aakanksha Rana and Cagri Ozcinar Email: ozcinarc at scss.tcd.ie
In the last few years, the data uploaded and exchanged between people has increased tremendously. Every minute Snapchat records 0.5M photo shares, people watch 4M videos on YouTube, people tweet 0.4M tweets and people share 50K pictures on Instagram. In all these examples the interesting fact is that all the data is unstructured and a lot of the volume comes from everyday media.
Digital surveillance is important to suppress the sharing of inappropriate content such as hateful videos, extremely violent videos or anything that can misguide viewers. This leads us to the next piece in our puzzle which is the availability of technology and resources to do this real-time.
We have made significant achievements in the field of image recognition through AI techniques, but their use is limited by resource-intensive algorithms or models.
In this project, we are going to develop deep neural models that would be much smaller than their predecessors. This would increase the number of devices these models can be used with such as mobile devices and other smaller IoT devices. The compact model developed in this project can be deployed on mobile devices which social media apps. This model can be used to censor the recorded content on the users’ device itself without it ever reaching the cloud reducing computation time, storage requirement at the server as well as the communications costs without drastically impacting user experience with the apps.
References
https://arxiv.org/pdf/1805.08168.pdf
https://arxiv.org/pdf/1710.09282.pdf
https://arxiv.org/pdf/1902.04103.pdf
Requirement:
Basic understanding of Deep-learning,
Strong Python programming skills with knowledge of Pytorch/TensorFlow tools
The ideal candidate should have an interest in model compression and deep learning algorithms in general and must have the ability to learn new tools and knowledge.