PR-190: A Baseline For Detecting Misclassified and Out-of-Distribution Examples In Neural Networks
- 1. A Baseline For Detecting Misclassified and Out-of-Distribution Examples In Neural Networks PR-190 Kang, MinGuk mingukkang1994@gmail.com Sep., 01, 2019 https://arxiv.org/abs/1610.02136
- 2. Unprecedented Successes(Motivation) Image Classification Real-World Applications https://arxiv.org/abs/1905.11946
- 4. Why Deep Neural Networks tend to be overconfident? ① Softmax Probabilities are computed with the fast-growing exponential function But… No Experimental Analysis https://arxiv.org/pdf/1706.04599.pdf Expected Calibration Error(ECE) ① Depth ↑ ② Filters ↑ ③ Batch Normalization 有 ④ Weight Decay ↓ It remains future work to understand why these trends affect calibration while improving accuracy.
- 5. Contributions of this Paper 1. They show the prediction probability of incorrect and out-of-distribution examples tends to be lower than the prediction probability for correct examples. 2. These prediction probabilities form our detection baseline, and we demonstrate its efficacy through various computer vision, natural language processing, and automatic speech recognition tasks. 3. They contribute one method which outperforms the baseline on some (but not all) tasks. 4. the designation of standard tasks and evaluation metrics for assessing the automatic detection of errors and out-of-distribution examples.
- 6. Evaluation Metrics In-distribution Fish: 99 Out-of-distribution Fish: 1 Cheating Neural Network: 99% accuracy! So, Accuracy is not appropriate metric for out-of-distribution detection.
- 7. Evaluation Metrics ① AUROC(Area Under Receiver Operating Characteristic Curve) ① AUPR(Area Under Precision Recall Curve) FPR(False Positive Rate): 𝐹𝑃 𝐹𝑃+𝑇𝑁 TPR(True Positive Rate): 𝑇𝑃 𝑇𝑃+𝐹𝑁 interpreted as the probability that a positive example has a greater detector score/value than a negative example (Fawcett, 2005). AUROC is not ideal when the positive class and negative class have greatly differing base rates Precision: 𝑇𝑃 𝑇𝑃+𝐹𝑃 Recall: 𝑇𝑃 𝑇𝑃+𝐹𝑁 interpreted as the probability that a positive example has a greater detector score/value than a negative example (Fawcett, 2005). AUROC is not ideal when the positive class and negative class have greatly differing base rates
- 8. Experiments(Misclassified?) Confi: 0.81 0.91 0.84 0.91 0.85 0.75 0.90 0.88 Average: 0.86 Confi: 0.90 0.95 0.85 0.95 0.92 0.88 0.95 0.86 Predict: 7 Actual: 7 Predict: 8 Actual: 8 Predict: 8 Actual: 8 Predict: 8 Actual: 8 Predict: 5 Actual: 5 Predict: 7 Actual: 7 Predict: 9 Actual: 9 Predict: 6 Actual: 6 Average: 0.91
- 9. Experiments(Out of Distribution) Wide(40-4) Prediction Train CIFAR10 Dataset Test Select Maximum Softmax Probability and Use it as out-of-distribution score
- 10. Experiments(NLP) Same Phenomenon was discovered in the NLP! Sentiment Classification Text Categorization Automatic Speech Recognition Experimental Results of Sentiment Classification
- 11. Improved Method Abnormality Module 1. Train a normal classifier and append an auxiliary decoder which reconstructs the input with in-distribution dataset. 2. Froze the blue layer. 3. Train red layers on clean and noised training examples. Finally the sigmoid output of the red layers scores how normal the input is
- 12. Improved Method Abnormality Module Abnormality Module is useful to detect out-of-distribution samples!
- 13. Expected Calibration Error(ECE) ① Depth ↑ ② Filters ↑ ③ Batch Normalization 有 ④ Weight Decay ↓ It remains future work to understand why these trends affect calibration while improving accuracy. On Calibration of Modern Neural Networks (2017.06.14) (2016.10.07)
- 14. A Simple Unified Framework for Detecting Out-of- Distribution Samples And Adversarial Attacks (2018.07.10) Training Confidence-Calibrated Classifiers for detecting Out-of-Distribution samples (2017.11.26) Train Generative Adversarial Networks to generate Boundary Samples. Class(k) Probability 1/k
- 15. Deep Anomaly Detection with Outlier Exposure (2018.12.11) Utilize Realistic Outliers instead of boundary samples Class(k) Probability 1/k In-distribution dataset Out-of-Distribution dataset Thank You!