This paper presents a baseline method for detecting misclassified and out-of-distribution examples in neural networks. The method uses the maximum softmax probability from the classifier's output to distinguish between correctly and incorrectly classified examples, as well as between in-distribution and out-of-distribution examples. The softmax probabilities are found to be higher for correctly classified examples, making them a useful indicator for detecting errors or abnormal examples. However, the softmax probabilities can be misleading when viewed in isolation, as they do not directly reflect the model's confidence. The paper evaluates the method across several tasks in computer vision, natural language processing, and automatic speech recognition. It shows that the softmax-based baseline is effective in detecting misclassified and out-of-distribution examples in these tasks. However, the method is not always the best, and the paper also presents an alternative method that uses an auxiliary decoder to reconstruct the input and assess the abnormality of the example. This method can sometimes outperform the softmax-based baseline. The paper also discusses the importance of evaluating detection methods using appropriate metrics, such as the area under the receiver operating characteristic (AUROC) and the area under the precision-recall (AUPR) curves. These metrics are used to assess the performance of the detection methods in distinguishing between in-distribution and out-of-distribution examples. The paper also highlights the need for further research in this area, as there is still room for improvement in detecting misclassified and out-of-distribution examples. The paper concludes that while softmax probabilities are not directly useful as confidence estimates, they can still provide a surprisingly effective way to detect whether an example is misclassified or from a different distribution than the training data. This creates a strong baseline for detecting errors and out-of-distribution examples, which the paper hopes future research will surpass.This paper presents a baseline method for detecting misclassified and out-of-distribution examples in neural networks. The method uses the maximum softmax probability from the classifier's output to distinguish between correctly and incorrectly classified examples, as well as between in-distribution and out-of-distribution examples. The softmax probabilities are found to be higher for correctly classified examples, making them a useful indicator for detecting errors or abnormal examples. However, the softmax probabilities can be misleading when viewed in isolation, as they do not directly reflect the model's confidence. The paper evaluates the method across several tasks in computer vision, natural language processing, and automatic speech recognition. It shows that the softmax-based baseline is effective in detecting misclassified and out-of-distribution examples in these tasks. However, the method is not always the best, and the paper also presents an alternative method that uses an auxiliary decoder to reconstruct the input and assess the abnormality of the example. This method can sometimes outperform the softmax-based baseline. The paper also discusses the importance of evaluating detection methods using appropriate metrics, such as the area under the receiver operating characteristic (AUROC) and the area under the precision-recall (AUPR) curves. These metrics are used to assess the performance of the detection methods in distinguishing between in-distribution and out-of-distribution examples. The paper also highlights the need for further research in this area, as there is still room for improvement in detecting misclassified and out-of-distribution examples. The paper concludes that while softmax probabilities are not directly useful as confidence estimates, they can still provide a surprisingly effective way to detect whether an example is misclassified or from a different distribution than the training data. This creates a strong baseline for detecting errors and out-of-distribution examples, which the paper hopes future research will surpass.