Resiliência do classificador de n-upla WiSARD

Abstract

Just as machine learning technologies are advancing and gaining more space in modern society, it becomes more important to warrant these do not fail without warning. Considering that today’s hardware is prone to failure or error under certain under certain conditions, execution errors as well as errors due to environmental conditions, it becomes important to analyse the impact these errors might have over an already implemented model. Considering this, in this work we seek to analyse the effects of different types of error over the WiSARD weightless neural network model’s performance. This work is focussed on bit-flip errors in different parts of the WiSARD network, with special attention to errors in the model’s trained parameters and how this kind of error compares with others. For fixed network hyperparameters and preprocessing procedures, the accuracy drop of the model is taken for seven different datasets, considering different error injection rates and error patterns.