Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
E(n) Equivariant Topological Neural Networks
Authors: Claudio Battiloro, Ege Karaismailoglu, Mauricio Tec, George Dasoulas, Michelle Audirac, Francesca Dominici
ICLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The broad applicability of ETNNs is demonstrated through two tasks of vastly different scales: i) molecular property prediction on the QM9 benchmark and ii) land-use regression for hyper-local estimation of air pollution with multi-resolution irregular geospatial data. The results indicate that ETNNs are an effective tool for learning from diverse types of richly structured data, as they match or surpass Sot A equivariant TDL models with a significantly smaller computational burden, thus highlighting the benefits of a principled geometric inductive bias. Our implementation of ETNNs can be found here. 5 EXPERIMENTS |
| Researcher Affiliation | Academia | 1Harvard University, 2ETH Zurich |
| Pseudocode | No | The paper describes the methodology using mathematical formulations and textual explanations within sections like 'E(N) EQUIVARIANT TOPOLOGICAL NEURAL NETWORKS' and 'ARCHITECTURE DESCRIPTION', but it does not include any explicitly labeled pseudocode blocks or algorithms. |
| Open Source Code | Yes | Our implementation of ETNNs can be found here. The code, data splits, and virtual environment needed to replicate the experiments and easily use the ETNN framework are provided at the following repository: https://github.com/NSAPH-Projects/topological-equivariant-networks. |
| Open Datasets | Yes | We evaluate our model on molecular property prediction using the QM9 dataset (Ramakrishnan et al., 2014), a comprehensive collection of quantum chemical calculations for small organic molecules. The prediction targets of PM2.5 measurements are obtained from a publicly available dataset (Wang et al., 2023) consisting of measurements by mobile air sensors installed on cars, corresponding to the last quarter of 2021 in the Bronx, NY, USA. PLUTO, NYC Dept. of Planning (New York City Department of City Planning, 2024) AADT, NYC Dept. of Transportation (New York City Department of Transportation, 2024) Open Street Map (OSMnx) (Boeing, 2017) Wei et al. (Wei et al., 2023) |
| Dataset Splits | Yes | For the molecular property prediction task: We used the same train/validation/test splits as introduced in EGNN(Satorras et al., 2021). For the air pollution downscaling task: we randomly select 70% of census tracts for training and split the remaining 30% equally in test and validation tracts. |
| Hardware Specification | Yes | All experiments that produced the reported configurations in both tasks were conducted using Nvidia A100-SXM4-40GB GPUs. Each system has a capacity of 80GB RAM, 40GB GPU memory, and 6 CPUs per unit. |
| Software Dependencies | No | The paper mentions the use of 'osmnx Python package (Boeing, 2017)' and refers to a 'virtual environment' in the reproducibility statement. However, it does not provide specific version numbers for key software components such as Python, PyTorch, or CUDA libraries used for the experiments, apart from the citation for osmnx. |
| Experiment Setup | Yes | Table 8: Hyperparameters for ETNN Model in the Molecular Task. This table lists specific values for Optimizer (Adam), Initial Learning Rate ([5e-4, 1e-3]), Learning Rate Scheduler (Cosine Annealing), Weight Decay (1e-5), Batch Size (96), Epochs ([100, 200, 350, 1000]), Number of Message Passing Layers ([4, 7, 10]), Hidden Units per Layer ([70, 104, 128, 182]), Activation Function (SiLU), Invariant Normalization ([True, False]), and Gradient Clipping ([True, False]). Table 10: Hyperparameters for ETNN Model in the Air Pollution Downscaling Task. This table lists specific values for Optimizer (Adam), Initial Learning Rate ([1e-3, 1e-2]), Learning Rate Scheduler (Cosine Annealing), Weight Decay (1e-4), Batch Size (1), Epochs (500), Number of Message Passing Layers (4), Hidden Units per Layer ([4, 32]), Activation Function (SiLU), Invariant Normalization (False), Gradient Clipping (True), and Dropout ([0.025, 0.25]). |