Efficient Generalized Spherical CNNs

Authors: Oliver Cobb, Christopher G. R. Wallis, Augustine N. Mavor-Parker, Augustin Marignier, Matthew A. Price, Mayeul d'Avezac, Jason McEwen

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate state-of-the-art performance on all spherical benchmark problems considered, both in terms of accuracy and parameter efficiency.
Researcher Affiliation Industry Oliver J. Cobb, Christopher G. R. Wallis, Augustine N. Mavor-Parker, Augustin Marignier, Matthew A. Price, Mayeul d Avezac & Jason D. Mc Ewen Kagenova Limited, Guildford GU5 9LD, UK
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code No The paper mentions 'fourpi AI 1 code' which links to a commercial product page (https://www.kagenova.com/products/fourpi AI/) and a 'private tensossht 4 code' that is 'Available on request from https://www.kagenova.com/', neither of which constitute publicly available open-source code for the described methodology.
Open Datasets Yes We consider the now standard benchmark problem of classifying MNIST digits projected onto the sphere. ... using the QM7 dataset (Blum & Reymond, 2009; Rupp et al., 2012). ... the SHREC 17 (Savva et al., 2017) competition dataset, containing 51k 3D object meshes.
Dataset Splits No For the spherical component we again adopt the convolutional architecture shown in Figure 4 except with one fewer efficient generalized layer. We use bandlimits of p L0, L1, L2, L3, L4q p10, 6, 6, 3, 1q, channels of p K0, K1, K2, K3, K4q p5, 16, 24, 32, 40q and τmax 6. One minor difference is that this time we include a skip connection between the ℓ 0 components of the fourth and fifth layer. We proceed the convolutional layers with two dense layers of size p256, 64q and use batch normalization between each layer.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU models, CPU types, or memory) used to run the experiments.
Software Dependencies No The paper mentions using 'ssht' and 'so3' software packages, and a 'Tensor Flow implementation', but it does not specify version numbers for these software dependencies, which is required for reproducibility.
Experiment Setup Yes We train the network for 50 epochs on batches of size 32, using the Adam optimizer (Kingma & Ba, 2015) with a decaying learning rate starting at 0.001. For the restricted generalized convolutions we follow the approach of Kondor et al. (2018) by using L1 regularization (regularization strength 10 5)...