[B] [M] DIET: from obese data to sparse face models

Master Assignment

DIET: from obese data to sparse face models

Type: Master EE/CS

Period: TBD

Student: (Unassigned)

If you are interested please contact :

Description:

The availability of large datasets and processing power has led to an enormous advance in artificial intelligence during the last decade. The sheer increase of number of parameters and FLOPS also has sparked research in trimming down/pruning of models while keeping (some) control over the accuracy. A development that has gained popularity are sparse neural networks, in which many connections normally present between two layers on a neural network have been removed.

It is interesting to study various sparse architectures and their performance in a face verification scenario.

Many airports have implemented an automatic border control system. A traveler presents its biometric passport containing facial biometric information stored in a template, the system takes a facial image, compares it to the template and if they match sufficiently according to an algorithm, the system allows access.

This is one of many verification examples where a biometric template is compared to a current sensor measurement of a fingerprint, face or iris. The template storage, sensor and matching algorithm are three independent components. The EU GDP Regulation implies strict requirements on the secure storage and safe use of biometric templates. One solution is to integrate the template storage, sensor and algorithm in a single tamperproof personal device. This device communicates the verification result to a central system without exposing biometric information. An example of such device is a debit/credit bank (smart) cards with an integrated fingerprint sensor, see for example  https://www.fingerprints.com/solutions/payments/.

Since sparse architectures require less space, it becomes an interesting candidate for a feature extraction and verification algorithm on a resource limited device such as a MCU or FPGA.

Assignment