Rotation and Translation Invariant Representation Learning with Implicit Neural Representations

Authors: Sehyun Kwon, Joo Young Choi, Ernest K. Ryu

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4. Experiments, Table 2. Clustering on semantics representation z., Table 3. Using IRL-INR as pretext task for SCAN outperformed other combinations using TARGET-VAE and Sim CLR., Table 4. Increasing latent dimension d of IRL-INR leads to better clustering performance.
Researcher Affiliation Academia 1Interdisciplinary Program in Articifial Intelligence, Seoul National University 2Department of Mathematical Sciences, Seoul National University.
Pseudocode No The paper describes its methodology in text but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code: https://github.com/ sehyunkwon/IRL-INR.
Open Datasets Yes MNIST(U) is derived from MNIST with random rotations and translations respectively sampled from Uniform([0, 2π)) and N(0, 52)., WM811k is a dataset of silicon wafer maps classified into 9 defect patterns (Wu et al., 2015)., 5HDB consists of 20,000 simulated projections of integrin αIIb in complex with integrin β3 (Lin et al., 2015; Bepler et al., 2019)..., d Sprites consists of 2D shapes procedurally generated from 6 ground truth independent latent factors (Higgins et al., 2017)., WHOI-Plankton is an expert-labeled image dataset for plankton (Orenstein et al., 2015)., Galaxy Zoo consists of 61,578 RGB color images of galaxies from the Sloan Digital Sky Survey (Lintott et al., 2008).
Dataset Splits No The paper explicitly defines training and test sets with specific counts (e.g., "7350 training set and 3557 test set images" for WM811k, "16,000 training set and 4000 test set images" for 5HDB, "50,000 training set and 11,578 test set images" for Galaxy Zoo, "1000 training set and 200 test set images" for WHOI-Plankton). However, a separate validation set with specific counts or percentages is not explicitly mentioned.
Hardware Specification Yes We run all our experiments on a single NVIDIA RTX 3090 Ti GPU with 24 GB memory.
Software Dependencies No The paper mentions several software components and architectures like ResNet18, MLP, Adam optimizer, Sim CLR, Torchvision, but it does not provide specific version numbers for these software libraries or frameworks (e.g., PyTorch 1.x or Python 3.x).
Experiment Setup Yes We use the Adam optimizer with learning rate 1 10 4, weight decay 5 10 4, and batch size 128. For the loss function scaling coefficients, we use λrecon = λconsis = 1 and λsymm = 15. We train the model for 200, 500, 2000, 100 epochs for MNIST(U), WM811k, WHOI-Plankton, and {5HDB, d Sprites, Galaxy zoo} respectively. (From Section 4.1 and Table 5 in Appendix B)