University of Twente Student Theses
As of Friday, 8 August 2025, the current Student Theses repository is no longer available for thesis uploads. A new Student Theses repository will be available starting Friday, 15 August 2025.
Self Organised Criticality and Avalanche Detection in Neural Networks
Akin, Tayfun (2025) Self Organised Criticality and Avalanche Detection in Neural Networks.
|
PDF
1MB |
| Abstract: | In the domain of Machine Learning, which deals with algorithms that are designed to extract patterns from data that is fed to them, Neural Networks (NN) are a fundamental tool for tackling complex problems that may require higher levels of pattern recognition [22]. They require a large amount of data to be trained as compared to other methods and are relatively expensive to both train and operate. However, their ability to learn tasks that are mainly associated with human capabilities and depending on the context, being able to outperform human cognition makes them the choice of method for some domains [29]. NNs are what would be considered a Black-box system, yielding minimum information about the purpose and the structure of the neurons within them [25]. By their design, the influence that the neurons have on each other will form according to the training data and in an unpredictable fashion. This leads to neurons forming groups among themselves to achieve a particular step in the processing of its input, where their purpose in the overall system will not be immediately explainable to an outside observer. The Black-box nature of the NNs causes them to be tough to evaluate for reliability and faults. This secretive nature means that certain components of the decision process (for cases where the decisions need multiple stages of reasoning to achieve) can not be tested individually, only the system as a whole. That means that the whole system can not be theoretically proven to be reliable and, only be assessed as far as the testing data allows. This fact raises large concerns for applications that are safety-critical, such as the automotive and the health industries [27]. One of the factors that hinder the reliability of NNs is the naturally occurring state of criticality [8]. This state increases the likelihood of the system to change its state with very minimal outside influence, making the system highly sensitive. Having the neurons set up in such a way that would result in critical behaviour, they get a chance to get the rest of the neurons to fire in a cascading manner, with their numbers of activations following a power-law. Naturally, when this kind of behaviour is not designed into the system, such disturbances can have negative effects on their performance. We aim to reduce the number of these disturbances in a system by analyzing the underlying structure which may lead to such behaviors and potentially devising a method in which the structure can be modified to minimize the effect of them. |
| Item Type: | Essay (Master) |
| Faculty: | EEMCS: Electrical Engineering, Mathematics and Computer Science |
| Subject: | 54 computer science |
| Programme: | Computer Science MSc (60300) |
| Link to this item: | https://purl.utwente.nl/essays/107564 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page