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 and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Existence, Stability and Scalability of Orthogonal Convolutional Neural Networks
Authors: El Mehdi Achour, François Malgouyres, Franck Mamalet
JMLR 2022 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The theoretical results are confirmed with experiments and the landscape of the regularization term is studied. Experiments on real data sets show that when orthogonality is used to enforce robustness, the parameter multiplying the regularization term can be used to tune a tradeoff between accuracy and orthogonality, for the benefit of both accuracy and robustness. |
| Researcher Affiliation | Collaboration | El Mehdi Achour EMAIL Institut de Mathématiques de Toulouse ; UMR 5219 Université de Toulouse ; CNRS UPS IMT F-31062 Toulouse Cedex 9, France François Malgouyres EMAIL Institut de Mathématiques de Toulouse ; UMR 5219 Université de Toulouse ; CNRS UPS IMT F-31062 Toulouse Cedex 9, France Franck Mamalet EMAIL Institut de Recherche Technologique Saint Exupéry, Toulouse, France |
| Pseudocode | Yes | Algorithm 1 Computing the list of singular values of K, when S = 1, (Sedghi et al., 2018). Input: kernel tensor: K RM C k k, channel size: N k Output: list of the singular values of K: σ 1: procedure COMPUTESINGULARVALUES(K,N) 2: transforms = np.fft.fft2(K, (N,N), axes=[0, 1]) np stands for numpy 3: sigma = np.linalg.svd(transforms, compute_uv=False) 4: return sigma 5: end procedure |
| Open Source Code | Yes | The code will be made available in the DEEL.LIP 4 library. 4 https://github.com/deel-ai/deel-lip |
| Open Datasets | Yes | Finally, we illustrate on Cifar10 and Imagenette data sets how the regularization parameter can be chosen to control the tradeoff between accuracy and orthogonality, for the benefit of both accuracy and robustness (see Section 3.2). ... Figure 5: Cifar10: Mean evolution of metrics ... Figure 6: Imagenette: Mean evolution of metrics ... Imagenette data set (Howard, 2020). The latter is a 10-class subset of Imagenet data set (Deng et al., 2009)... |
| Dataset Splits | No | The optimized network achieves 91% accuracy on the Cifar10 test set, for the Conv2D classical network. Acc. clean: Classical accuracy on a clean test set Erob : The empirical robustness accuracy, i.e. the proportion of test samples on which a vanilla Projected Gradient Descent (PGD) attack (Madry et al., 2018) failed (for a robustness radius ϵ = 36/255). The paper mentions a "test set" and "test samples" but does not specify the ratios or sizes of train/validation/test splits. |
| Hardware Specification | No | No hardware details are specified in the paper. |
| Software Dependencies | No | transforms = np.fft.fft2(K, (N,N), axes=[0, 1]) np stands for numpy sigma = np.linalg.svd(transforms, compute_uv=False) Adam optimizer (Kingma and Ba, 2015) The paper mentions numpy and Adam optimizer, but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | We train a single 2D convolutional layer with circular padding. We explore all the architectures such that K 2 = , for C J1, 64K , M J1, 64K, S {1, 2, 4}, and k {1, 3, 5, 7}. ... For each architecture, the model is trained using a Glorot uniform initializer and an Adam optimizer (Kingma and Ba, 2015) with fixed learning rate 0.01 on a null loss (Ltask(X, Y, K) = 0, for all input X, target Y, and kernel tensor K) and the Lorth(K) regularization (see Definition 1) during 3000 steps. ... Batch size (which thus has no influence on the results) is set to one. ... The network is trained during 400 epochs with a batch size of 128, using cross-entropy loss with temperature, Adam optimizer (Kingma and Ba, 2015) with a decreasing learning rate, and standard data augmentation. ... The initial learning rate is set to 0.03, and linearly decreased down to 3 10 4. |