It is expected that the demand for larger and larger data rates will continue to rise in the coming years with applications like smart sensor networks, distributed machine learning, digital twins of production facilities, etc. With the recent advances in machine learning techniques, semantic communication has been proposed a solution to push beyond the boundaries of conventional communication systems, as semantic communication does not require the accurate recovery of all information available at the sender, but rather the recovery of the meaning that the sender wants to convey with a specific task in mind. 

As a example, the concept of semantic communication can outperform classical communication methods in a scenario, where each agent in a multi agent network observes some data with the goal of solving a task together. As a more specific example, in [1] each agent has one part of an image available, and the goal at the central receiver is to classify the whole image correctly. In the classical approach the image would be transmitted in its entirety, and the classification algorithm would use the received image as its input. However, it was shown that the data rate can be greatly reduced by using the semantic communication approach, where only the meaning of the content of each image part is transmitted and the receiver uses only the received semantic information to carry out the classification.

To design such a semantic communication system, the optimization problem is usually formulated to maximize the mutual information between the received symbols at the decoder and the semantic variable. However, as the direct optimization of the mutual information is intractable as the optimal decoder is not available, a lower bound is usually optimized instead. 

The key metric for the semantic communication is whether a task got executed correctly. Therefore, to compare the semantic communication approach to the classical approach, a task execution error rate needs to be calculated. For the example of image classification, the semantic decoder directly outputs the most likely image class. For the classical communication approach, the image is reconstructed at the receiver. Then the reconstructed image is put into a separate image classifier, where it is then evaluated if the classification is correct or not. However, in current simulation setups, this separate image classifier is usually trained on the original data and not the distorted reconstructed images. It is an open question whether the classification accuracy can be improved if the separate image classifier is trained on the reconstructed images instead of the original data.

[1] Beck E, Bockelmann C, Dekorsy A. Semantic Information Recovery in Wireless Networks. Sensors. 2023; 23(14):6347. https://doi.org/10.3390/s23146347

Objectives:

    • Develop and implement a simulation model of a semantic encoder decoder structure using deep neural networks 

    • Implement a simulation model of a classical joint source channel coding encoder decoder structure using deep neural networks for image reconstruction

    • Investigate the performance of the classification accuracy: 

        ◦ Is the classification accuracy improved if the classifier was trained on the reconstructed image of the classical communication system compared to the case when the classifier was trained on the original data 

        ◦ Investigate if simply adding Gaussian noise to the training dataset can increase the classification accuracy of the  reconstructed image of the classical communication system

    • Perform numeric evaluations of the developed simulation model: different SNR, number of senders, different datasets, etc.

Requirements:

    • Information theory background knowledge

    • Basic knowledge of machine learning 

    • Programming skills (python)