Topological Neural Networks go Persistent, Equivariant, and Continuous
Authors: Yogesh Verma, Amauri H Souza, Vikas Garg
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, (continuous and E(n)-equivariant extensions of) Top Nets achieve strong performance across diverse tasks, including antibody design, molecular dynamics simulation, and drug property prediction. |
| Researcher Affiliation | Collaboration | 1Department of Computer Science, Aalto University, Finland 2Federal Institute of Ceará 3Yai Yai Ltd. |
| Pseudocode | No | The paper describes algorithms and steps in prose and mathematical equations but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available here: https://github.com/Aalto-QuML/TopNets |
| Open Datasets | Yes | Tasks. We assess the performance of Top Nets on diverse tasks: (i) we evaluate our method performance on real-world graph classification data while considering discrete and continuous versions of various GNNs and TNNs in Section 6.1, (ii) we benchmark Top Nets efficacy in property prediction using QM9 molecular data, highlighting the effectiveness of its equivariant variant in Section 6.2, (iii) we demonstrate Top Nets utility in co-designing antibody sequence and structure using the SAb Dab database in Section 6.3, and (iv) we evaluate our method on 3BPA MD17 trajectories (Kovács et al., 2021) in Section 6.4. |
| Dataset Splits | Yes | Following the data preparation strategy of Eijkelboom et al. (2023); Satorras et al. (2021), we partition the dataset into training, validation, and test sets. |
| Hardware Specification | Yes | We conducted an ablation study to characterize the runtime complexity of our method, assessing the time taken per epoch to train different models on a single V100 GPU. |
| Software Dependencies | No | The paper mentions using 'odeint package' but does not specify its version number or any other software dependencies with version details. |
| Experiment Setup | Yes | Table 9: Default hyperparameters for Top Nets for Graph Classification Benchmark. Hyperparameter Meaning Value Solver ODE-Solver adaptive-heun,euler GNN GNN Architecture {GCN,GIN,MPSN} PH Type of PH {VC,TOGL,Re PHINE} Steps Number of steps for ODE solver {20,15,10,5} Node Hidden Dim Latent dimension of node features 128 PH embed dim Latent dimension of PH features 64 Num Filt Number of filtrations 8 Hiden Filtration Hidden dimension of filtration functions 16 Batch Size Size of batches 64 LR Learning Rate 0.001 Scheduler Learning Rate scheduler Cosine-Annealing-LR Epochs Number of epochs 300 |