Learning Partial Equivariances From Data
Authors: David W. Romero, Suhas Lohit
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate Partial G-CNNs on illustrative toy tasks and vision benchmark datasets. We show that whenever full equivariance is beneficial, e.g., for rotated MNIST, Partial G-CNNs learn to remain fully equivariant. However, if equivariance becomes harmful, e.g., for classification of 6 / 9 digits and natural images, Partial G-CNNs learn to adjust equivariance to a subset of the group to improve accuracy. Partial G-CNNs improve upon conventional G-CNNs when equivariance reductions are advantageous, and match their performance whenever their design is optimal. |
| Researcher Affiliation | Collaboration | David W. Romero Vrije Universiteit Amsterdam Amsterdam, The Netherlands d.w.romeroguzman@vu.nl Suhas Lohit Mitsubishi Electric Research Laboratories Cambridge, MA, USA slohit@merl.com |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | 2Our code is publicly available at github.com/merlresearch/partial_gcnn. |
| Open Datasets | Yes | MNIST dataset [31]...Rot MNIST [30], CIFAR-10 and CIFAR-100 [28]. |
| Dataset Splits | No | The paper uses datasets like Rot MNIST, CIFAR-10, and CIFAR-100 but does not specify the train/validation/test splits (e.g., percentages or counts) or refer to standard predefined splits for these specific experiments. For MNIST6-180 and MNIST6-M, it describes how they were constructed but not how they were split for training/validation/testing. |
| Hardware Specification | No | The paper mentions being carried out on 'the Dutch national infrastructure with the support of SURF Cooperative' but does not provide specific hardware details such as GPU/CPU models, memory, or processor types used for the experiments. |
| Software Dependencies | No | The paper mentions the use of 'Py Torch convolution primitives' but does not specify version numbers for PyTorch or any other software dependencies, such as Python version, specific libraries, or CUDA versions. |
| Experiment Setup | Yes | Experimental details. We parameterize all our convolutional kernels as 3-layer SIRENs [41] with 32 hidden units. All our networks except for the (partial) group equivariant 13-layer CNNs [29] used in Sec. 5.1 are constructed with 2 residual blocks of 32 channels each, batch normalization [24] following the structure shown in Fig. 3. ... Additional experimental details such as specific hyperparameters used and complementary results can be found in Appx. E, F. ... we utilize a 10x lower learning rate for the parameters of the probability distributions (See Appx. E.3 for details). |