Learning Color Equivariant Representations
Authors: Yulong Yang, Felix O'Mahony, Christine Allen-Blanchette
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the utility of our approach on synthetic and real world datasets where we consistently outperform competitive baselines. 5 EXPERIMENTS In this section, we highlight the sample efficiency of our model in the context of hue generalization; the stability of our hue-equivariant representations compared to CEConv (Lengyel et al., 2024); the utility of a notion of color equivariance that includes saturation and luminance; and the utility of our representations for color-based sorting. Our models achieve strong performance on extensive experiments, against competitive baselines. |
| Researcher Affiliation | Academia | Yulong Yang1, , Felix O Mahony2, Christine Allen-Blanchette1, 1Princeton University, Princeton, USA 2University of Oxford, Oxford, UK EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper defines group actions and group convolution mathematically but does not contain explicit pseudocode or algorithm blocks with structured steps. |
| Open Source Code | Yes | Our source code is publicly available online at https://github.com/CAB-Lab-Princeton/Learning-Color-Equivariant-Representations. |
| Open Datasets | Yes | Our dataset is a variation of MNIST (Le Cun et al., 1998) with 60k training examples and 10k test examples classified into one of 10 categories. We demonstrate improved generalization to local hue-shifts and significantly reduced equivariance error compared to CEConv on the 3D Shapes classification dataset (Burgess & Kim, 2018). We demonstrate improved generalization to saturation-shifts compared to Res Net50 (He et al., 2016) and CEConv-3 on the Camelyon17 classification dataset (Bandi et al., 2018). We demonstrate improved generalization to global color-shifts compared to Res Net (He et al., 2016) and CEConv on the Caltech-101 (Li et al., 2022), Oxford-IIT Pets (Parkhi et al.), Stanford Cars (Krause et al., 2013), STL-10 (Coates et al., 2011), CIFAR-10 and CIFAR-100 (Krizhevsky et al., 2009) datasets. We include classification results on the Image Net dataset (Deng et al., 2009). |
| Dataset Splits | Yes | Our dataset is a variation of MNIST (Le Cun et al., 1998) with 60k training examples and 10k test examples classified into one of 10 categories. The 3D Shapes dataset consists of RGB images of 3D shapes... There are 48k examples in the training set, and 12k examples in each of the test sets. The Camelyon17 dataset consists of images of human tissue... The dataset consists of 387,490 examples, 302,436 of which are used for training and 85,054 of which are used for testing. CIFAR-10... The dataset consists of 60k examples, 50k of which are used for training and 10k for testing. CIFAR-100... The dataset consists of 60k examples, 40k of which are used for training, 10k for validation, and 10k for testing. Caltech-101 (Li et al., 2022) classification dataset. The dataset consists of 9,146 examples, 2/3 of which are used for training and 1/3 for testing. ... The training set consists of 6941 examples, the validation set consists of 868 examples, and the test set consists of 868 examples. STL-10 (Coates et al., 2011) classification dataset. The dataset consists of 5k training examples and 8k testing examples. Stanford Cars (Krause et al., 2013) classification dataset. ... The dataset consists of 8,144 training examples and 8,041 testing examples. Oxford-IIT Pets (Parkhi et al.) classification dataset. The dataset consists of 3,680 training examples and 3,669 testing examples. Image Net (Deng et al., 2009) classification dataset. The dataset consists 1000 categories with 14,197,122 annotated images. |
| Hardware Specification | Yes | All models except Image Net were trained on a shared research computing cluster. Each compute node allocates an Nvidia L40 GPU, 24 core partitions of an Intel Xeon Gold 5320 CPU, and 24GBs of DDR4 3200MHz RDIMMs. Image Net was trained on compute nodes with 8 Nvidia L40 GPUs, two Intel Xeon Gold 5320 CPUs, and 512GBs of DDR4 3200MHz RDIMMs. |
| Software Dependencies | No | The paper mentions using Adam optimizer (Kingma & Ba, 2014) and SGD, but does not specify software versions for programming languages, libraries, or other dependencies. |
| Experiment Setup | Yes | We train our equivariant networks and the conventional architectures for 5 epochs with a batch size of 128. We optimize over a cross-entropy loss using the Adam optimizer (Kingma & Ba, 2014) with β1 = 0.9, and β2 = 0.999. We use an initial learning rate of 10 3 for the Z2 network, and 10 4 for the hue-equivariant network. We train the Res Net architectures for 100k iterations with a batch size of 128. We optimize over a cross-entropy loss using SGD with a learning rate of 10 2. We train our equivariant networks and the conventional architectures for 10k iterations with a batch size of 32. We optimize over a cross-entropy loss using the Adam optimizer (Kingma & Ba, 2014) with an initial learning rate of 10 2, β1 = 0.9, and β2 = 0.999. We train our equivariant networks and the conventional architectures for 300 epochs, and a batch size of 128. We optimize over a cross-entropy loss using SGD with an initial learning rate of 10 1 and a cosine-annealing scheduler. |