Weisfeiler Leman for Euclidean Equivariant Machine Learning

Authors: Snir Hordan, Tal Amir, Nadav Dym

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically evaluate our claims via a separation experiment on combinatorial graphs that are 2-WL distinguishable, yet are 1-WL indistinguishable, via the EXP (Abboud et al., 2020) dataset.
Researcher Affiliation Academia 1Faculty of Mathematics, Technion Israel Institute of Technology, Haifa, Israel 2Faculty of Computer Science, Technion Israel Institute of Technology, Haifa, Israel.
Pseudocode Yes We define the Convolution Layer of We LNet , i.e. We LConv, which is based on EGNN, as c(i, j) = PPGNan(θ; , T)(X, V )i,j (10) mij = ϕe(hi, hj, eij, c(i, j)) (11) mi = j mi,j (12) xout i = xi + ϕn(mi)vi + X j ϕx(mi,j)(xj xi) + X j =i ϕv(mi,j)vj (13) vout i = ˆ ϕn(mi)vi + X j ˆϕx(mi,j)(xj xi) + X j =i ˆϕv(mi,j)vj (14) hout i = ϕh(hi, mi) (15)
Open Source Code Yes Code is available at https://www.github.com/ Intelli Finder/welnet
Open Datasets Yes We empirically evaluate our claims via a separation experiment on combinatorial graphs that are 2-WL distinguishable, yet are 1-WL indistinguishable, via the EXP (Abboud et al., 2020) dataset.
Dataset Splits Yes Following (Victor Garcia Satorras, 2021), we sampled 3,000 trajectories for training, 2,000 for validation and 2.000 for testing.
Hardware Specification Yes We ran our experiment on an NVIDIA A40 GPU with CUDA toolkit version 12.1.
Software Dependencies Yes We ran our experiment on an NVIDIA A40 GPU with CUDA toolkit version 12.1.
Experiment Setup Yes Table 4. Configuration of the We LNet Architecture. HYPERPARAMETER VALUE ACTIVATION SCALED SOFTPLUS EDGE FEATURES DIM 128 WL FEATURES DIM 32 LEARNING RATE 1E-3 OPTIMIZER ADAM SCHEDULER STEPLR NUMBER OF CONVOLUTIONS 4 2-WL ITERATIONS (T) 2