ReLUs Are Sufficient for Learning Implicit Neural Representations

Authors: Joseph Shenouda, Yamin Zhou, Robert D Nowak

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we demonstrate that, contrary to popular belief, one can learn state-of-the-art INRs based on a DNN composed of only Re LU neurons. We substantiate our claims through experiments in signal representation, super resolution, and computed tomography, demonstrating the versatility and effectiveness of our method.
Researcher Affiliation Academia 1Department of Electrical and Computer Engineering, University of Wisconsin-Madison 2Department of Computer Sciences, University of Wisconsin-Madison. Correspondence to: Joseph Shenouda <jshenouda@wisc.edu>.
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes The code for all experiments can be found at https://github. com/joeshenouda/relu-inrs.
Open Datasets Yes In this experiment we simulated CT reconstruction by taking 100 equally spaced CT measurements of a 326 × 435 chest X-ray image (Clark et al., 2013). In this experiment we fit the four INR architectures to the standard 256 × 256 cameraman image. We implemented 4× super resolution on a single image from the DIV2K image dataset (Agustsson & Timofte, 2017).
Dataset Splits No The paper mentions training details like epochs and learning rates but does not explicitly provide training, validation, or test dataset split percentages or counts.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., CPU, GPU models, or memory) used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions).
Experiment Setup Yes For all experiments we utilized a three hidden layer DNN. The full training details for all experiments can be found in Appendix D. For the BW-Re LU DNN we used an initial learning rate of η0 = 4e-3 a scaling parameter applied to each neuron of c = 9. All models were trained for 1000 epochs and the decay rate for the learning rate was r = 0.1. The architecture consisted of 3 hidden layers and 300 neurons per layer.