Deep Neural Networks (DNNs) are known as black box algorithmsthat lack transparency and interpretability for humans. eXplainableArtificial Intelligence (XAI) is introduced to tackle this problem. MostXAI methods are utilized post-training, providing explanations of themodel to clarify its predictions and inner workings for human understanding. However, there is a shortage of methods that utilize XAIduring training to not only observe the model’s behavior but alsoexploit this information for the benefit of the model.In our approach, we propose a novel method that leverages XAIduring the training process itself. Incorporating feedback from XAIcan give us insights into important features of input data that impact model decisions. This work explores focusing more on specificfeatures during training, which could potentially improve model performance introspectively throughout the training phase. We analyzethe stability of feature explanations during training and find thatthe model’s attention to specific features is consistent in the MNISTdataset. However, unimportant features lack stability. The OCTMNIST dataset, on the other hand, has stable explanations for important features but less consistent explanations for less significant features. Based on this observation, two types of masks, namely fixedand dynamic, are applied to the model’s structure using XAI’s feedback with minimal human intervention. These masks identify themore important features from the less important ones and set the pixels associated with less significant features to zero. The fixed mask isgenerated based on XAI feedback after the model is fully trained, andthen it is applied to the output of the first convolutional layer of a newmodel (with the same architecture), which is trained from scratch. Onthe other hand, the dynamic mask is generated based on XAI feedback during training, and it is applied to the model while the modelis still training. As a result, these masks are changing during different epochs. Examining these two methods on both deep and shallowmodels, we find that both masking methods, particularly the fixedone, reduce the focus of all models on the least important parts of theinput data. This results in improved accuracy and loss in all models.As a result, this approach enhances the model’s interpretability andperformance by incorporating XAI into the training process.
| Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:hh-51728 |
| Date | January 2023 |
| Creators | Dastkarvelayati, Rozhin, Ghafourian, Soudabeh |
| Publisher | Högskolan i Halmstad, Akademin för informationsteknologi |
| Source Sets | DiVA Archive at Upsalla University |
| Language | English |
| Detected Language | English |
| Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
| Format | application/pdf |
| Rights | info:eu-repo/semantics/openAccess |
Page generated in 0.0023 seconds