Equivariant Polynomial Functional Networks
Authors: Thieu Vo, Hoang V. Tran, Tho Tran Huu, An Nguyen The, Thanh Tran, Minh-Khoi Nguyen-Nhat, Duy-Tung Pham, Tan Minh Nguyen
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate MAGEP-NFNs on three tasks: predicting CNN generalization from weights using Small CNN Zoo (Unterthiner et al., 2020), weight space style editing, and classifying INRs using INRs data (Zhou et al., 2024b). Experimental results show that our model achieves competitive performance and efficiency compared to existing baselines. |
| Researcher Affiliation | Collaboration | 1National University of Singapore 2FPT Software AI Center, Vietnam 3Vin University, Vietnam. |
| Pseudocode | Yes | H.1. Equivariant Layers Pseudocode H.1.1. PSEUDOCODE FOR CASE i = L H.1.2. PSEUDOCODE FOR CASE i = 1 H.1.3. PSEUDOCODE FOR CASE 1 < i < L H.2. Invariant Layers Pseudocode |
| Open Source Code | No | The paper mentions utilizing official code for baselines (available at: https://github.com/Allan Yang Zhou/nfn), but does not provide specific code release for the methodology described in this paper (MAGEP-NFN). |
| Open Datasets | Yes | We utilize the Small CNN Zoo dataset (Unterthiner et al., 2020), which contains various pretrained CNN models. Following (Tran et al., 2024a), we employ three distinct INR weight datasets (Zhou et al., 2024b), each was trained on a different image dataset: CIFAR-10 (Krizhevsky & Hinton, 2009), Fashion-MNIST (Xiao et al., 2017), and MNIST (Le Cun & Cortes, 2005). |
| Dataset Splits | Yes | The Tanh subset from the CNN Zoo dataset has 5,949 training instances and 1,488 testing instances, while the original Re LU subset consists of 6,050 training instances and 1,513 testing instances. Table 6: Dataset Train size Val size Original Re LU 6050 1513 Augment Re LU 12100 3026 Tanh 5949 1488 Table 9: Dataset size for Classifying INRs task. CIFAR-10 Train 45000 Val 5000 Test 10000 MNIST size Train 45000 Val 5000 Test 10000 Fashion-MNIST 45000 Val 5000 Test 20000 |
| Hardware Specification | Yes | The entire training process on an A100 GPU takes 30 minutes. Approximately 2 hours on an A100 GPU. Approximately 35 minutes on an A100 GPU. |
| Software Dependencies | No | The paper does not explicitly list specific software dependencies with version numbers. It mentions 'einops-style pseudocode' in the Appendix, implying its use, but no version is provided. |
| Experiment Setup | Yes | Table 8: Hyperparameters for MAGEP-NFN on prediciting generalization task. MLP hidden 500 Loss Binary cross-entropy Optimizer Adam Learning rate 0.001 Batch size 8 Epoch 50. Table 10: Hyperparameters of MAGEP-NFN for each dataset in Classify INRs task. MNIST MAGEP-NFN hidden dimension 128x3 Base model MAGEP-Inv Base model hidden dimension 128 MLP hidden neurons 1000 Dropout 0.1 Learning rate 0.0001 Batch size 32 Step 200000 Loss Binary cross-entropy. Table 13: Hyperparameters for MAGEP-NFN on weight space style editing task. MAGEP-NFN hidden dimension 16 NP dimension 128 Optimizer Adam Learning rate 0.001 Batch size 32 Steps 50000. |