A Unified Algebraic Perspective on Lipschitz Neural Networks

Authors: Alexandre Araujo, Aaron J Havens, Blaise Delattre, Alexandre Allauzen, Bin Hu

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, the comprehensive set of experiments on image classification shows that SLLs outperform previous approaches on certified robust accuracy.
Researcher Affiliation Collaboration 1 INRIA, Ecole Normale Sup erieure, CNRS, PSL University, Paris, France 2 CSL & ECE, University of Illinois Urbana-Champaign, IL, USA 3 Miles Team, LAMSADE, Universit e Paris-Dauphine, PSL University, Paris, France 4 Foxstream, Vaulx-en-Velin, France 5 ESPCI PSL, Paris, France
Pseudocode No No pseudocode or algorithm blocks were found in the paper.
Open Source Code Yes Code is available at github.com/araujoalexandre/Lipschitz-SLL-Networks.
Open Datasets Yes First, we evaluate our networks (SLL) on CIFAR10 and CIAFR100 and compare the results against recent 1-Lipschitz neural network structures: Cayley, SOC, SOC+, CPL and AOL. We have also implemented SLL on Tiny Image Net (see Table 2).
Dataset Splits Yes First, we evaluate our networks (SLL) on CIFAR10 and CIAFR100 and compare the results against recent 1-Lipschitz neural network structures: Cayley, SOC, SOC+, CPL and AOL. We have also implemented SLL on Tiny Image Net (see Table 2).
Hardware Specification No The paper mentions '4 GPUs' in Table 4 and 'HPC resources from GENCI IDRIS' in the Acknowledgments. However, no specific GPU models (e.g., NVIDIA A100, Tesla V100), CPU models, or detailed hardware configurations were provided for reproducibility.
Software Dependencies No The paper mentions using an 'Adam optimizer (Kingma et al., 2014)' and 'Cross Entropy loss as in Prach et al. (2022)'. However, it does not specify version numbers for general software dependencies such as Python, PyTorch, TensorFlow, CUDA, or specific library versions used for implementation.
Experiment Setup Yes We trained our networks with a batch size of 256 over 1000 epochs with the data augmentation used by . We use an Adam optimizer (Kingma et al., 2014) with 0.01 learning rate and parameters β1 and β2 equal to 0.5 and 0.9 respectively and no weight decay. We use a piecewise triangular learning rate scheduler to decay the learning rate during training. We use the Cross Entropy loss as in Prach et al. (2022) with a temperature of 0.25 and an offset value 3