Beyond the Single Neuron Convex Barrier for Neural Network Certification
Authors: Gagandeep Singh, Rupanshu Ganvir, Markus Püschel, Martin Vechev
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experimental results show that k-Re LU enables significantly more precise certification than existing state-of-the-art verifiers while maintaining scalability. |
| Researcher Affiliation | Academia | Gagandeep Singh1, Rupanshu Ganvir2, Markus Püschel1, Martin Vechev1 Department of Computer Science ETH Zurich, Switzerland 1{gsingh,pueschel,martin.vechev}@inf.ethz.ch 2rganvir@student.ethz.ch |
| Pseudocode | No | No explicitly labeled pseudocode or algorithm blocks found. |
| Open Source Code | Yes | We made k Poly publicly available as part of the ERAN [28] framework for neural network verification. [28] ERAN: ETH Robustness Analyzer for Neural Networks, 2018. [Online]. Available: https://github.com/eth-sri/eran |
| Open Datasets | Yes | Neural networks We used 9 MNIST [31] and CIFAR10 [32] fully connected (FNNs), convolutional (CNNs), and residual networks with Re LU activations shown in Table 2. |
| Dataset Splits | No | No explicit mention of specific training, validation, or test dataset splits (e.g., percentages or sample counts) for a validation set. |
| Hardware Specification | Yes | The runtimes of all experiments for the MNIST FNNs were measured on a 3.3 GHz 10 Core Intel i9-7900X Skylake CPU with a main memory of 64 GB whereas the experiments for the rest were run on a 2.6 GHz 14 core Intel Xeon CPU E5-2690 with 512 GB of main memory. |
| Software Dependencies | No | k Poly is written in Python and uses cdd [25, 26] for computing convex hulls, and Gurobi [27] for refining Deep Poly Re LU relaxations and proving that ψ holds with k-Re LU relaxations. The paper does not provide specific version numbers for these software components. |
| Experiment Setup | Yes | The last column of Table 2 shows the value of k for all networks. For the MNIST and CIFAR10 Conv Big networks, we encode the first 3 Re LU layers with 1-Re LU while the remaining are encoded with 5-Re LU. We use l = 3 in (7) for encoding 5-Re LU. For the remaining 3 networks, we encode the first Re LU layer with 1-Re LU while the remaining layers are encoded adaptively. Here, we choose a value of k for which the total number of calls to 3-Re LU is 500. |