Adaptive Verifiable Training Using Pairwise Class Similarity

نویسندگان

چکیده

Verifiable training has shown success in creating neural networks that are provably robust to a given amount of noise. However, despite only enforcing single robustness criterion, its performance scales poorly with dataset complexity. On CIFAR10, non-robust LeNet model 21.63% error rate, while created using verifiable and L-infinity criterion 8/255, an rate 57.10%. Upon examination, we find when labeling visually similar classes, the model's is as high 61.65%. Thus, attribute loss inter-class similarity. Classes (i.e., close feature space) increase difficulty learning model. While it may be desirable train for large region, pairwise class similarities limit potential gains. Furthermore, consideration must made regarding relative cost mistaking one another. In security or safety critical tasks, classes likely belong same group, thus equally sensitive. this work, propose new approach utilizes similarity improve create models respect multiple adversarial criteria. First, cluster agglomerate clustering assign criteria based on degree between clusters. Next, two methods apply our approach: (1) Inter-Group Robustness Prioritization method, which uses custom term guarantees (2) decision tree trains sub-classifiers different combines them architecture. Our experiments Fashion-MNIST CIFAR10 demonstrate by prioritizing most dissimilar groups, clean up 9.63% 30.89% respectively. CIFAR100, reduces 26.32%.

برای دانلود رایگان متن کامل این مقاله و بیش از 32 میلیون مقاله دیگر ابتدا ثبت نام کنید

اگر عضو سایت هستید لطفا وارد حساب کاربری خود شوید

منابع مشابه

Face Recognition Using Balanced Pairwise Classifier Training

This paper presents a novel pairwise classification framework for face recognition (FR). In the framework, a two-class (intraand inter-personal) classification problem is considered and features are extracted using pairs of images. This approach makes it possible to incorporate prior knowledge through the selection of training image pairs and facilitates the application of the framework to tack...

متن کامل

Learning Pairwise Similarity

Overview Different vectorial and / or (dis)similarity representations can be produced for a given data. These distinct representations or data generating models have typically been used individually, in single classifiers or single clustering algorithms, or simultaneously, as in classifier combination techniques or cluster ensemble methods, depending, respectively, on whether working under a su...

متن کامل

Beyond Pairwise Shape Similarity Analysis

This paper considers two major applications of shape matching algorithms: (a) query-by-example, i. e. retrieving the most similar shapes from a database and (b) finding clusters of shapes, each represented by a single prototype. Our approach goes beyond pairwise shape similarity analysis by considering the underlying structure of the shape manifold, which is estimated from the shape similarity ...

متن کامل

Pairwise similarity scores using functional annotations: review and comparison

Previous works have already checked the relationship of a number of functional similarity scores with different types of biological information: sequences, protein families, expression profiles, pathways, literature, etc. This work aims to provide a review and a critical comparison of previously proposed similarity scores, to help in the decision of choosing the appropriate one for the problem ...

متن کامل

Classifying Microarray Data Using Pairwise Similarity between Gene Profiles

Figure 1: Dendrogram with a threshold applied (---) to create clusters. The first cluster has genes A and B while the second one has genes C, D, and E. genes to be measured simultaneously. However, interpretation of the large data sets is complex. Commonly used methods include hierarchical clustering and k-nearest-neighbor [3]. After applying these methods, clusters of co-expressed genes are pr...

متن کامل

ذخیره در منابع من


  با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید

ژورنال

عنوان ژورنال: Proceedings of the ... AAAI Conference on Artificial Intelligence

سال: 2021

ISSN: ['2159-5399', '2374-3468']

DOI: https://doi.org/10.1609/aaai.v35i11.17223