Geometrical aspects of lattice gauge equivariant convolutional neural networks
Authors: David I. Müller, Jimmy Aronsson, Daniel Schuh
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Theoretical | In this paper, we revisit L-CNNs from a geometric point of view and extend them by including a larger degree of lattice symmetry. First, we review lattice gauge theory and the original formulation of L-CNNs in Section 2. L-CNNs were originally constructed by incorporating local symmetry into ordinary CNNs, which means that L-CNNs are equivariant under lattice translations but not under other lattice symmetries such as rotations and reflections. We remedy this in Section 3 by applying methods from G-CNNs to LCNNs. Our main result is a gauge equivariant convolution that can be applied to tensor fields and that is equivariant under translations, rotations, and reflections. Finally, in Section 4, we put the original L-CNNs in a broader context by relating them to a mathematical theory for equivariant neural networks. In doing so, we demonstrate how convolutions in L-CNNs can be understood as discretizations of convolutions on SU(N) principal bundles. |
| Researcher Affiliation | Academia | David I. Müller EMAIL TU Wien, Institute for Theoretical Physics, A-1040 Vienna, Austria Jimmy Aronsson EMAIL Chalmers University of Technology, Department of Mathematical Sciences, SE-412 96 Gothenburg, Sweden Daniel Schuh EMAIL TU Wien, Institute for Theoretical Physics, A-1040 Vienna, Austria |
| Pseudocode | No | The paper describes methods and operations using mathematical equations and textual explanations, but it does not include any explicitly labeled pseudocode or algorithm blocks. For instance, the descriptions of convolutional layers (Eq. 17), bilinear layers (Eq. 22), and activation functions (Eq. 23) are presented as mathematical formulas. |
| Open Source Code | No | The paper does not contain any explicit statement about releasing source code for the methodology described, nor does it provide a link to a code repository. It mentions Py Torch (Paszke et al., 2019) as a framework that could simplify implementations, but this is a general tool and not a release of their own specific code. |
| Open Datasets | No | The paper primarily focuses on theoretical aspects and extensions of L-CNNs. It discusses their application to lattice gauge theories and mentions that 'L-CNNs can accurately learn gauge invariant observables such as Wilson loops from datasets of gauge field configurations' (page 2), but it does not specify or provide access information for any particular dataset used in the context of this paper's contributions. No experiments using specific datasets are conducted. |
| Dataset Splits | No | The paper is theoretical and does not conduct experiments involving datasets. Therefore, there is no mention of dataset splits for training, validation, or testing. |
| Hardware Specification | No | The paper mentions 'available memory on modern GPUs' and discusses the 'memory footprint of G-equivariant L-CNNs' (page 14) in the context of computational requirements of the proposed models. However, it does not specify any particular hardware (GPU/CPU models, etc.) used for running experiments, as the paper does not report on empirical experiments. |
| Software Dependencies | No | The paper mentions 'Py Torch (Paszke et al., 2019)' as a framework where implementation could be simplified (page 5). However, it does not specify version numbers for PyTorch or any other software component that would be needed to replicate any experimental results, as the paper is theoretical and does not present experiments. |
| Experiment Setup | No | The paper is theoretical and does not describe any experimental setup, hyperparameter values, training configurations, or system-level settings. It focuses on the mathematical formulation and extension of L-CNNs. |