Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]
Explaining, Fast and Slow: Abstraction and Refinement of Provable Explanations
Authors: Shahaf Bassan, Yizhak Yisrael Elboher, Tobias Ladner, Matthias Althoff, Guy Katz
ICML 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments demonstrate that our approach enhances the efficiency of obtaining provably sufficient explanations for neural network predictions while additionally providing a fine-grained interpretation of the network s predictions across different abstraction levels. We evaluate our algorithm on various benchmarks and show that it significantly outperforms existing verification-based methods by producing much smaller explanations and doing so substantially more efficiently. |
| Researcher Affiliation | Academia | 1Hebrew University of Jerusalem 2Technical University of Munich. Correspondence to: <EMAIL>, <EMAIL>, <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 Minimal Explanation Search Algorithm 2 Minimal Abstract Explanation Search |
| Open Source Code | No | The paper does not explicitly state that the authors' implementation code for the methodology is open-source or provide a link to a code repository. It mentions using 'CORA (Althoff, 2015) as the backend neural network verifier', which is a third-party tool. |
| Open Datasets | Yes | We performed our experiments on three image classification benchmarks: (a) MNIST (Le Cun, 1998), (b) CIFAR-10 (Krizhevsky & Hinton, 2009), and (c) GTSRB (Stallkamp et al., 2012). |
| Dataset Splits | Yes | For MNIST and CIFAR-10, we utilized common models from the neural network verification competition (VNN-COMP) (Brix et al., 2023), which are frequently used in experiments related to neural network verification tasks. Specifically, the MNIST model architecture is sourced from the ERAN benchmark within VNN-COMP, and the CIFAR-10 model is derived from the marabou benchmark. Since GTSRB is not directly utilized in VNN-COMP, we trained this model using a batch size of 32 for 10 epochs with the ADAM optimizer, achieving an accuracy of 84.8%. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU, GPU models, memory) used for running the experiments. It only mentions implementing algorithms using a backend verifier and training models. |
| Software Dependencies | No | The paper mentions 'CORA (Althoff, 2015) as the backend neural network verifier' and 'ADAM optimizer' but does not specify version numbers for these or any other key software libraries or programming languages. |
| Experiment Setup | Yes | Since GTSRB is not directly utilized in VNN-COMP, we trained this model using a batch size of 32 for 10 epochs with the ADAM optimizer, achieving an accuracy of 84.8%. The precise dimensions and configurations of the models used for both VNN-COMP (MNIST and CIFAR-10) and GTSRB are provided: Tab. 3 for MNIST, Tab. 4, and Tab. 5 for GTSRB. For MNIST and GTSRB, we use a perturbation radius ϵ = 0.01 as commonly used in VNN-COMP benchmarks, and for CIFAR-10, we use a smaller perturbation radius ϵ = 0.001 as we have found this network to be not very robust. For simplicity, we start with a coarsest abstraction at a reduction rate ρ = 10% of the original network s neurons and increase ρ by 10% at each subsequent refinement, until ρ = 100% is reached, which restores the original network. |