This research article proposes a method to understand and interpret the classification decisions of automated image classification systems, particularly focusing on non-linear classifiers. The authors introduce a methodology called *layer-wise relevance propagation* that allows for pixel-wise decomposition of predictions in kernel-based classifiers and multilayer neural networks. This method visualizes the contributions of individual pixels to the final prediction as heatmaps, enabling human experts to verify the system's reasoning and identify regions of interest. The approach is evaluated on various datasets, including PASCAL VOC 2009 images, synthetic image data, MNIST handwritten digits, and the pre-trained ImageNet model. The paper also discusses an alternative method based on *Taylor decomposition* and compares it with layer-wise relevance propagation, showing that the latter can be implemented without approximation for a wide range of architectures. The authors highlight the importance of this work in making machine learning models more interpretable and transparent.This research article proposes a method to understand and interpret the classification decisions of automated image classification systems, particularly focusing on non-linear classifiers. The authors introduce a methodology called *layer-wise relevance propagation* that allows for pixel-wise decomposition of predictions in kernel-based classifiers and multilayer neural networks. This method visualizes the contributions of individual pixels to the final prediction as heatmaps, enabling human experts to verify the system's reasoning and identify regions of interest. The approach is evaluated on various datasets, including PASCAL VOC 2009 images, synthetic image data, MNIST handwritten digits, and the pre-trained ImageNet model. The paper also discusses an alternative method based on *Taylor decomposition* and compares it with layer-wise relevance propagation, showing that the latter can be implemented without approximation for a wide range of architectures. The authors highlight the importance of this work in making machine learning models more interpretable and transparent.