Provable Editing of Deep Neural Networks using Parametric Linear Relaxation

Authors: Zhe Tao, Aditya V Thakur

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate PREPARED i) using the VNN-COMP benchmarks, ii) by editing CIFAR10 and TINYIMAGENET image-recognition DNNs, and BERT sentiment-classification DNNs for local robustness, and iii) by training a DNN to model a geodynamics process and satisfy physics constraints.
Researcher Affiliation Academia Zhe Tao University of California, Davis Davis, CA 95616, USA zhetao@ucdavis.edu Aditya V. Thakur University of California, Davis Davis, CA 95616, USA avthakur@ucdavis.edu
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks. The methods are described using mathematical definitions and textual explanations.
Open Source Code No For the implementation of our tool, we could not make the code open access due to ongoing IP restrictions.
Open Datasets Yes We evaluate PREPARED i) using the VNN-COMP benchmarks, ii) by editing CIFAR10 and TINYIMAGENET image-recognition DNNs, and BERT sentiment-classification DNNs for local robustness, and iii) by training a DNN to model a geodynamics process and satisfy physics constraints. ... The CIFAR10 DNN has 17.2M parameters... The Tiny Image Net DNN has 51.9M parameters... (citing [21] for CIFAR10, [33] for TINYIMAGENET, and [29] for VNN-COMP 22 benchmarks).
Dataset Splits Yes The validation set is the same 10% of the training set, randomly selected with seed 0. ... We train 2000 epochs and take the epoch with the highest validation accuracy.
Hardware Specification Yes All experiments were run on a machine with Dual Intel Xeon Silver 4216 Processor 16-Core 2.1GHz with 384 GB of memory, SSD and RTX-A6000 with 48 GB of GPU memory running Ubuntu 20.04.
Software Dependencies No We have implemented PREPARED in Py Torch [32] and use Gurobi [17] as the LP solver. ... We use verifiers α,β-CROWN [46, 44], MN-Ba B [9], or Deep T [4], depending on the task. ... We use auto_Li RPA [50] or Deep Poly [39] to compute the constant bounds for the input to the first layer to edit. (No explicit version numbers stated directly with the software names).
Experiment Setup Yes Appendix D.2, D.3, D.4, and D.5 provide detailed hyperparameters for each experiment, including optimizers, learning rates, batch sizes, number of epochs, and layer freezing strategies.