фотография книги онлайн бесплатно и без регистрации
Главное меню:
Интересное
фотография книги онлайн бесплатно и без регистрации, Evaluating the Robustness of Neural Networks: An Extreme Value , STAIR: Improving Safety Alignment with Introspective Reasoning, Clever: A Curated Benchmark for Formally Verified Code Generation, Counterfactual Debiasing for Fact Verification, Tractable Multi-Agent Reinforcement Learning through Behavioral , LLaVA-OneVision: Easy Visual Task Transfer | OpenReview, MIND over Body: Adaptive Thinking using Dynamic Computation.
Evaluating the Robustness of Neural Networks: An Extreme Value Theory Approach | OpenReviewToggle navigationOpenReview.net×Go to ICLR 2018 Conference homepageEvaluating the Robustness of Neural Networks: An Extreme Value Theory ApproachTsui-Wei Weng*, Huan Zhang*, Pin-Yu Chen, Jinfeng Yi, Dong Su, Yupeng Gao, Cho-Jui Hsieh, Luca Daniel 15 Feb 2018 (modified: 22 Jun 2025)ICLR 2018 Conference Blind SubmissionReaders: EveryoneAbstract: The robustness of neural networks to adversarial examples has received great attention due to security implications. Despite various attack approaches to crafting visually imperceptible adversarial examples, little has been developed towards a comprehensive measure of robustness. In this paper, we provide theoretical justification for converting robustness analysis into a local Lipschitz constant estimation problem, and propose to use the Extreme Value Theory for efficient evaluation. Our analysis yields a novel robustness metric called CLEVER, which is short for Cross Lipschitz Extreme Value for nEtwork Robustness. The proposed CLEVER score is attack-agnostic and is computationally feasible for large neural networks. Experimental results on various networks, including ResNet, Inception-v3 and MobileNet, show that (i) CLEVER is aligned with the robustness indication measured by the $\ell_2$ and $\ell_\infty$ norms of adversarial examples from powerful attacks, and (ii) defended networks using defensive distillation or bounded ReLU indeed give better CLEVER scores. To the best of our knowledge, CLEVER is the first attack-independent robustness metric that can be applied to any neural network classifiers.TL;DR: We propose the first attack-independent robustness metric, a.k.a CLEVER, that can be applied to any neural network classifier.Keywords: robustness, adversarial machine learning, neural network, extreme value theory, adversarial example, adversarial perturbationCode: [ huanzhang12/CLEVER](https://github.com/huanzhang12/CLEVER)Community Implementations: [ 2 code implementations](https://www.catalyzex.com/paper/evaluating-the-robustness-of-neural-networks/code)13 RepliesLoadingJoin the TeamJoin the Team is a long-term project to advance science through improved peer review with legal nonprofit status. We gratefully acknowledge the support of the . © 2025 OpenReview, Our analysis yields a novel robustness metric called CLEVER, which is short for Cross Lipschitz Extreme Value for nEtwork Robustness. The proposed CLEVER score is attack-agnostic and is computationally feasible for large neural networks., One common approach is training models to refuse unsafe queries, but this strategy can be vulnerable to clever prompts, often referred to as jailbreak attacks, which can trick the AI into providing harmful responses. Our method, STAIR (SafeTy Alignment with Introspective Reasoning), guides models to think more carefully before responding..
We introduce CLEVER, the first curated benchmark for evaluating the generation of specifications and formally verified code in Lean. The benchmark comprises of 161 programming problems; it evaluates both formal speci-fication generation and implementation synthesis from natural language, requiring formal correctness proofs for both., 579 In this paper, we have proposed a novel counter- factual framework CLEVER for debiasing fact- checking models. Unlike existing works, CLEVER is augmentation-free and mitigates biases on infer- ence stage. In CLEVER, the claim-evidence fusion model and the claim-only model are independently trained to capture the corresponding information., A significant roadblock to the development of principled multi-agent reinforcement learning (MARL) algorithms is the fact that desired solution concepts like Nash equilibria may be intractable to, We present LLaVA-OneVision, a family of open large multimodal models (LMMs) developed by consolidating our insights into data, models, and visual representations in the LLaVA-NeXT blog series. Our, Clever use of intermediate activations to assess input complexity. Should be able to work with existing architectures making engineering it for downstream real-world use cases simpler., Clever design: reformulating the ILP to a weighted bipartite matching / assignment problem and using Hungarian algorithm that has shorter solving time than communication time (so we can have actual speedup)..