The paper " Sanity Checks for Saliency Maps" by Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim addresses the challenge of evaluating the effectiveness and limitations of saliency methods in machine learning models. Saliency methods are popular tools used to highlight features in inputs that are relevant for model predictions. However, the authors argue that relying solely on visual assessment can be misleading. To address this issue, the authors propose a methodology based on randomization tests to evaluate the adequacy of explanation approaches. They apply these tests to several saliency methods for image classification with neural networks, finding that some methods are independent of both the model and the data generating process. Consequently, these methods are inadequate for tasks that depend on either the model or the data, such as debugging the model or explaining the relationship between inputs and outputs. The paper includes two main types of randomization tests: a model parameter randomization test and a data randomization test. The model parameter randomization test assesses the sensitivity of a saliency method to changes in the model's learned parameters, while the data randomization test evaluates the method's dependence on the relationship between instances and labels in the data. Through extensive experiments, the authors find that methods like Guided Backpropagation and its variants are insensitive to model parameters and data transformations, making them inadequate for tasks requiring model-specific or data-dependent explanations. They also show that visual inspection can be misleading, as some methods produce visually similar outputs to edge detectors, which do not depend on the model or training data. The paper concludes by discussing the implications of these findings and suggesting that visual inspection alone is insufficient for assessing the quality and relevance of saliency maps. The authors hope that their work will serve as a guide for researchers in designing more rigorous evaluation methods for new model explanations.The paper " Sanity Checks for Saliency Maps" by Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim addresses the challenge of evaluating the effectiveness and limitations of saliency methods in machine learning models. Saliency methods are popular tools used to highlight features in inputs that are relevant for model predictions. However, the authors argue that relying solely on visual assessment can be misleading. To address this issue, the authors propose a methodology based on randomization tests to evaluate the adequacy of explanation approaches. They apply these tests to several saliency methods for image classification with neural networks, finding that some methods are independent of both the model and the data generating process. Consequently, these methods are inadequate for tasks that depend on either the model or the data, such as debugging the model or explaining the relationship between inputs and outputs. The paper includes two main types of randomization tests: a model parameter randomization test and a data randomization test. The model parameter randomization test assesses the sensitivity of a saliency method to changes in the model's learned parameters, while the data randomization test evaluates the method's dependence on the relationship between instances and labels in the data. Through extensive experiments, the authors find that methods like Guided Backpropagation and its variants are insensitive to model parameters and data transformations, making them inadequate for tasks requiring model-specific or data-dependent explanations. They also show that visual inspection can be misleading, as some methods produce visually similar outputs to edge detectors, which do not depend on the model or training data. The paper concludes by discussing the implications of these findings and suggesting that visual inspection alone is insufficient for assessing the quality and relevance of saliency maps. The authors hope that their work will serve as a guide for researchers in designing more rigorous evaluation methods for new model explanations.