English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Zendy Plus

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Zendy Tools

Zendy Open

Deep Neural Network (DNN) models are widely used for image classi8 fication. While they offer high performance in terms of accuracy, researchers are 9 concerned about if these models inappropriately make inferences using features ir10 relevant to the target object in a given image. To address this concern, we propose 11 a metamorphic testing approach that assesses if a given inference is made based 12 on irrelevant features. Specifically, we propose two metamorphic relations (MRs) 13 to detect such unreliable inferences. These relations expect (a) the classification 14 results with different labels or the same labels but less certainty from models after 15 corrupting the relevant features of images, and (b) the classification results with 16 the same labels after corrupting irrelevant features. The inferences that violate the 17 metamorphic relations are regarded as unreliable inferences. 18 Our evaluation demonstrated that our approach can effectively identify unre19 liable inferences for single-label classification models with an average precision of 20 64.1% and 96.4% for the two MRs, respectively. As for multi-label classification 21 models, the corresponding precision for MR-1 and MR-2 is 78.2% and 86.5%, re22 spectively. Further, we conducted an empirical study to understand the problem 23 of unreliable inferences in practice. Specifically, we applied our approach to 18 24 pre-trained single-label image classification models and 3 multi-label classification 25 models, and then examined their inferences on the ImageNet and COCO datasets. 26 We found that unreliable inferences are pervasive. Specifically, for each model, 27 more than thousands of correct classifications are actually made using irrelevant 28 features. Next, we investigated the effect of such pervasive unreliable inferences, 29 and found that they can cause significant degradation of a model’s overall accuracy. 30 After including these unreliable inferences from the test set, the model’s accuracy 31 can be significantly changed. Therefore, we recommend that developers should pay 32 more attention to these unreliable inferences during the model evaluations. We also 33 explored the correlation between model accuracy and the size of unreliable infer34 ences. We found the inferences of the input with smaller objects are easier to be 35 unreliable. Lastly, we found that the current model training methodologies can 36 guide the models to learn object-relevant features to certain extent, but may not 37 necessarily prevent the model from making unreliable inferences. We encourage 38 the community to propose more effective training methodologies to address this 39 issue. 40

To what extent do DNN-based image classification models make unreliable inferences?