Difference between revisions of "Ternary Neural Networks for Face Recognition"
From iis-projects
(Created page with "==Introduction== Face recognition is among the many tasks that can be solved with great accuracy using deep learning techniques. Recent work has demonstrated that this task ca...") |
(No difference)
|
Revision as of 17:36, 16 July 2021
Contents
Introduction
Face recognition is among the many tasks that can be solved with great accuracy using deep learning techniques. Recent work has demonstrated that this task can be successfully tackled with networks using highly quantized (1-bit) weights, known as Binary Weight Networks. This suggests that other aggressively quantized network families may be successfully applied to the problem to increase the energy efficiency of the final system. One such class of networks are ternary neural networks (TNNs), where all weights and activations are quantized to 3 levels: {-1, 0, 1}.
Project description
In this project, you will train a ternary neural network (TNN) for face recognition on the "Labelled Faces in the Wild" (LFW) dataset, optimize it for accuracy and prepare it for deployment on CUTIE, our in-house TNN accelerator.
You will perform the following steps:
- Familiarize with the basics of face recognition
- Train a baseline full-precision network which can reach good accuracy on face recognition on the LFW dataset
- Ternarize and optimize this network
- Map the network to CUTIE and perform power simulations to determine the energy cost per inference
Required Skills
- Basic knowledge of deep learning
- Basic knowledge of Python
Skills you might find useful, but are not required:
- Previous experience with highly quantized neural networks
Literature
- [1] Labelled Faces in the Wild
- [2] SphereFace: Deep Hypersphere Embedding for Face Recognition
- [3] P. Jokic et al., Battery-Less Face Recognition at the Extreme Edge
Professor
Status: Available
Possible to complete as a Semester or Bachelor Thesis
Supervision: Georg Rutishauser georgr@iis.ee.ethz.ch, Moritz Scherer scheremo@iis.ee.ethz.ch
Meetings & Presentations
The students and advisor(s) agree on weekly meetings to discuss all relevant decisions and decide on how to proceed. Of course, additional meetings can be organized to address urgent issues. At the end of the project, you have to present/defend your work during a 15 min. presentation and 5 min. of discussion as part of the IIS colloquium.