Neural networks achieve state-of-the-art performance in many complex machine learning tasks (e.g., Object Detection, Image Classification, Audio Recognition etc.) In doing so, the respective models (weights and biases) size has exploded. This results in great power consumption, high inference latency and icreased memory complexity. It is therefore of interest to find ways to compress these models in order to achieve energy savings, inference speed and storing requirements. One very popular method to do so, is quantization. The task of the student is to explain how Post Training Quantization (PTQ) is applied, given that fixed-point representation is assumed [1].

[1] 2106.08295.pdf (arxiv.org)

It is nice for the student to have some background knowledge on deep learning, e.g., what is a neural network, how it is represented, how it is trained etc. However, introductory material can be provided if a student is eager to learn, and questions on topics that are unclear to the student are always welcome.

In general, this is a student-driven task, therefore it is the student's job to plan and execute the review of the given paper. Support and guidance will be gladly provided if requested. There will also be a clear discussion of what is required in the final presentation as well as evaluation points, directly after the topic assignment.

5G and beyond wireless communication systems gave/are expected to give rise to many new applications with stringent requirements that were not achievable with 4G or earlier systems. These requirements are: low latency, high reliability, battery life maximization among others.

To this extent, short blocklength coding has gained particular interest, and new research on efficient coding schemes is emerging. An idea previously discussed in the literature is the Ordered Statistics Decoder (OSD) [1]. Recent works in literature are addressing its strengths, its shortcomings and are suggesting ways to make this decoder more practical and efficient in the short blocklength regime.

The objective of the student is to study the seminal paper [1], that introduces the idea of ordered statistics decoding and be able to describe how the decoder works, what are its strengths, what are its limitations.

[1] Soft-decision decoding of linear block codes based on ordered statistics

Compared with traditional frame-based cameras, an event-based camera has the advantages of low latency, high dynamic range, (almost) no motion blur, etc., and can respond fast to a brightness change at the image plane with thresholds determined by the previous state of brightness. It can generate event data of structural features without signal processing, such as edges and corners, which saves time, energy and computing effort.

However, it still lacks standard processes for analyzing, characterizing and evaluating event-based camera information. Moreover, data acquisition of this camera demands changes in brightness on a respective pixel. This can be introduced artificially by relative movement of the camera to the object. This might miss out part of the data due to linear translation along structural elements like edges or introduce some error due to error motion or vibration.

This thesis aims to ensure safety and quality when implementing event-based cameras in the field of industrial inspection, by dedicated experiments and related discussion of results. Event based camera imageswill be characterized and evaluated. New methods for event generation and signal processing will be proposed which will make use of the special characteristics of event-based cameras and their special characteristics arising from their working principle.