Learned layer-level concepts in the machine learning model on the physical layer in 6G L1

Thumbnail Image

Files

master_Vu_Duy_2025.pdf (26.43 MB)

URL

Journal Title

Journal ISSN

Volume Title

School of Science | Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.

Authors

Date

2025-04-07

Department

Major/Subject

Machine Learning, Data Science and Artificial Intelligence

Mcode

Degree programme

Master's Programme in Computer, Communication and Information Sciences

Language

en

Pages

109

Series

Abstract

As the telecommunications industry evolves from 5G to 6G, the integration of artificial intelligence (AI) and deep learning models into communication networks is becoming increasingly vital. While AI has demonstrated significant success in this field, the complexity of their solutions comes at the expense of interpretability. Gaining insight into the inner workings of these models is crucial for building trust in stakeholders, enhancing transparency, aiding developers in improving performance, and facilitating the identification of potential issues. In light of this, this thesis will investigate the interpretability of deep learning-based receiver applications. We propose using Neural Activation Pattern (NAP) as an interpretability technique and evaluate its applicability in our use case. Since NAP technique consists of two distinct components, clustering inputs processed similarly by the neural network and identifying learned layer concepts, these aspects serve as the basis for evaluating NAP technique in our deep learning receiver. Additionally, we proposed modifications to the original technique to improve clustering results. Our enhanced approach incorporates a distribution estimation method and an alternative normalization method to address the limitations of the standard NAP technique. Experimental results show that the proposed modifications can extract a greater number of patterns that are more stable, particularly when employing kernel density estimation as the distribution estimation method and standardization as the normalization method. The computational cost of the enhanced method is significantly higher, which might be mitigated by parallelization, at the expense of resource usage, meaning there is still a need for further improving the overall effiency. In terms of interpretability, as far as signal-to-noise ratio (SNR) and velocity are concerned, our analysis reveals that the most prominent concepts are characterized by SNR values, whereas the velocity is not proven to be a key concept. These results are agreed by both the original NAP technique by Bäuerle et al. and the modified one, though in several layers, the statistics of SNR can be different between the two variants. Nonetheless, we believe that interpreting and understanding the learned concepts from NAP technique requires further research and exploration. For example, regarding robustness of clustering algorithm, as the data in the thesis is left-skewed, log-transformation or power transformation could be explored as an alternative method for standardization; whereas regarding clustering evaluation, the viability of Density-Based Clustering Validation as a better metric for comparing clustering solutions is also a possible research direction.

Description

Supervisor

Linna, Riku

Thesis advisor

Tuononen, Marko

Keywords

clustering, concept discovery, deep learning, explainable AI, machine learning, machine learning interpretability, physical layer, wireless communication

Other note

Citation

Permanent link to this item

https://urn.fi/URN:NBN:fi:aalto-202505203931

Collections

[dipl] Perustieteiden korkeakoulu / SCI