General Nonlinearities in SO(2)-Equivariant CNNs

Authors: Daniel Franzen, Michael Wand

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In experiments with 2D and 3D data, we obtain results that compare favorably to the state-of-the-art in terms of accuracy while permitting continuous symmetry and exact equivariance.
Researcher Affiliation Academia Daniel Franzen Institute of Computer Science Johannes Gutenberg University Mainz Staudingerweg 9, 55122 Mainz, Germany dfranz@uni-mainz.de Michael Wand Institute of Computer Science Johannes Gutenberg University Mainz Staudingerweg 9, 55122 Mainz, Germany wandm@uni-mainz.de
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes A more comprehensive presentation, along with source code, is provided in the supplementary material.
Open Datasets Yes We test our implementation on image and 3D data. In the image case, we replicate the architecture used by Weiler and Cesa [39] on the MNIST-rot dataset from their recent survey of E(2)-equivariant image processing networks. ... For the 3D surfel case, we use Model Net-40 [44] as benchmark.
Dataset Splits No The paper mentions training models with a batch size but does not provide specific details on validation data splits (percentages, counts, or explicit methodology for splitting).
Hardware Specification Yes In Table 1, we list training time in seconds per epoch on a single Nvidia RTX 2080 Ti graphics card.
Software Dependencies No The paper mentions 'Py Torch' and 'Py Ke Ops [2]' but does not provide specific version numbers for these software components.
Experiment Setup Yes Architecture: Our concrete network design follows the construction of Weiler and Cesa [39] for their best MNIST-rot models. We use the same number of equivariant and linear layers with the same output channel count and filter properties (radii, rotation orders and width) and also apply Dropout [33] with p = 0.3 before each linear layer. We train our models for 40 epochs with a batch size of 64 images. We use the Adam [21] optimizer, starting with a learning rate of 0.015, with an exponential decay factor of 0.8 per epoch starting after epoch 16.