The Hessian perspective into the Nature of Convolutional Neural Networks

Authors: Sidak Pal Singh, Thomas Hofmann, Bernhard Schölkopf

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

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically validate our upper bounds in a variety of settings, in particular, with both linear and Re LU activations, MSE and CE loss, as well as on datasets such as CIFAR10, Fashion MNIST, MNIST, and a synthetic dataset.
Researcher Affiliation Academia 1ETH Zürich, Switzerland 2MPI for Intelligent Systems, Tübingen, Germany.
Pseudocode No The paper describes mathematical proofs, theorems, and derivations but does not include any pseudocode or algorithm blocks.
Open Source Code Yes The corresponding code can be found under https://github.com/sidak/hessian_perspective_cnns.
Open Datasets Yes We empirically validate our upper bounds in a variety of settings, in particular, with both linear and Re LU activations, MSE and CE loss, as well as on datasets such as CIFAR10, Fashion MNIST, MNIST, and a synthetic dataset.
Dataset Splits No The paper mentions using datasets like CIFAR10, but it does not specify explicit training, validation, or test splits (e.g., 80/10/10 split percentages or sample counts for each set) needed for reproducibility.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., CPU, GPU models, memory) used to run the experiments. It only mentions general experimental settings.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies (e.g., Python, PyTorch, TensorFlow, specific solvers or libraries) used in their experiments, which is necessary for reproducibility.
Experiment Setup Yes We empirically validate our upper bounds in a variety of settings, in particular, with both linear and Re LU activations, MSE and CE loss, as well as on datasets such as CIFAR10, Fashion MNIST, MNIST, and a synthetic dataset. Following (Singh et al., 2021), to rigorously illustrate the match with our bounds, we compute the exact Hessians, without approximations, and in Float64 precision.