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].

ViViT: Curvature Access Through The Generalized Gauss-Newton’s Low-Rank Structure

Authors: Felix Dangel, Lukas Tatzel, Philipp Hennig

TMLR 2023 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate this by conducting performance benchmarks and substantiate Vi Vi T s usefulness by studying the impact of noise on the GGN s structural properties during neural network training.
Researcher Affiliation Academia Felix Dangel EMAIL University of Tübingen, Tübingen, Germany Lukas Tatzel EMAIL University of Tübingen, Tübingen, Germany Philipp Hennig EMAIL University of Tübingen & MPI for Intelligent Systems, Tübingen, Germany
Pseudocode No The paper describes methods and processes through mathematical equations and textual explanations, but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes We introduce approximations that allow a flexible trade-off between computational cost and accuracy, and provide a fully-featured efficient implementation in Py Torch (Paszke et al., 2019) on top of the Back PACK (Dangel et al., 2020) package at https://github.com/f-dangel/vivit. The code used for the experiments is available at https://github.com/f-dangel/vivit-experiments.
Open Datasets Yes Architectures include three deep convolutional neural networks from Deep OBS (Schneider et al., 2019) (2c2d on Fashion-MNIST, 3c3d on CIFAR-10 and All-CNN-C on CIFAR-100), as well as residual networks from He et al. (2016) on CIFAR-10 based on Idelbayev (2018) all are equipped with cross-entropy loss.
Dataset Splits Yes In experiments with fixed mini-batches the batch sizes correspond to Deep OBS default value for training where possible (CIFAR-10: N = 128, Fashion-MNIST: N = 128). The residual networks use a batch size of N = 128. On CIFAR-100 (trained with N = 256), we reduce the batch size to N = 64 to fit the exact computation on the full mini-batch, used as baseline, into memory. If the GGN approximation is evaluated on a subset of the mini-batch (sub), N/8 of the samples are used (as in Zhang et al. (2017)).
Hardware Specification Yes Results in this section were generated on a workstation with an Intel Core i7-8700K CPU (32 GB) and one NVIDIA Ge Force RTX 2080 Ti GPU (11 GB).
Software Dependencies No The paper mentions software like PyTorch and Back PACK and cites their original publications, but does not provide specific version numbers for these libraries used in the experiments.
Experiment Setup Yes We train the following Deep OBS (Schneider et al., 2019) architectures with SGD and Adam: 3c3d on CIFAR-10, 2c2d on Fashion-MNIST and All-CNN-C on CIFAR-100 all are equipped with cross-entropy loss. To ensure successful training, we use the hyperparameters from Dangel et al. (2020) (see Table S.3). We also train a residual network Res Net-32 He et al. (2016) with cross-entropy loss on CIFAR-10 with both SGD and Adam. For this, we use a batch size of 128 and train for 180 epochs. Momentum for SGD was fixed to 0.9, and Adam uses the default parameters (β1 = 0.9, β2 = 0.999, ϵ = 10 8).