A1 Journal article (refereed)
Node co-activations as a means of error detection : Towards fault-tolerant neural networks (2022)
Myllyaho, L., Nurminen, J. K., & Mikkonen, T. (2022). Node co-activations as a means of error detection : Towards fault-tolerant neural networks. Array, 15, Article 100201. https://doi.org/10.1016/j.array.2022.100201
JYU authors or editors
Publication details
All authors or editors: Myllyaho, Lalli; Nurminen, Jukka K.; Mikkonen, Tommi
Journal or series: Array
ISSN: 2590-0056
eISSN: 2590-0056
Publication year: 2022
Publication date: 10/06/2022
Volume: 15
Article number: 100201
Publisher: Elsevier
Publication country: Netherlands
Publication language: English
DOI: https://doi.org/10.1016/j.array.2022.100201
Publication open access: Openly available
Publication channel open access: Open Access channel
Publication is parallel published (JYX): https://jyx.jyu.fi/handle/123456789/82179
Abstract
Machine learning has proved an efficient tool, but the systems need tools to mitigate risks during runtime. One approach is fault tolerance: detecting and handling errors before they cause harm.
Objective:
This paper investigates whether rare co-activations – pairs of usually segregated nodes activating together – are indicative of problems in neural networks (NN). These could be used to detect concept drift and flagging untrustworthy predictions.
Method:
We trained four NNs. For each, we studied how often each pair of nodes activates together. In a separate test set, we counted how many rare co-activations occurred with each input, and grouped the inputs based on whether its classification was correct, incorrect, or whether its class was absent during training.
Results:
Rare co-activations are much more common in inputs from a class that was absent during training. Incorrectly classified inputs averaged a larger number of rare co-activations than correctly classified inputs, but the difference was smaller.
Conclusions:
As rare co-activations are more common in unprecedented inputs, they show potential for detecting concept drift. There is also some potential in detecting single inputs from untrained classes. The small difference between correctly and incorrectly predicted inputs is less promising and needs further research.
Keywords: machine learning; neural networks (information technology); reliability (general); errors
Free keywords: machine learning; fault tolerance; neural networks; error detection; concept drift; dependability
Contributing organizations
Ministry reporting: Yes
Reporting Year: 2022
JUFO rating: 1