$(\textrm{Implicit})^2$: Implicit Layers for Implicit Representations
Authors: Zhichun Huang, Shaojie Bai, J. Zico Kolter
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically evaluated our method on learning multiple implicit representations for images, audios, videos, and 3D models, showing that our (Implicit)2 approach substantially improve upon existing models while being both faster to train and much more memory efficient. Overall, our results of learning implicit representation on various domains (including images, audios, videos) suggest that the (Implicit)2 approach offers clear improvements over existing explicit models used for these purposes, where we are able to achieve the same or better level of performance while being up to 3 more memory-efficient and 3 more time-efficient. |
| Researcher Affiliation | Academia | Zhichun Huang Carnegie Mellon University Pittsburgh, PA 15213 zhichunh@cs.cmu.edu Shaojie Bai Carnegie Mellon University Pittsburgh, PA 15213 shaojieb@cs.cmu.edu J. Zico Kolter Carnegie Mellon University Pittsburgh, PA 15213 zkolter@cs.cmu.edu |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | We provide our implementation1 and the descriptions of the tasks and datasets can be found in the appendix. ^1Official implementation can be found at https://github.com/locuslab/Imp Sq |
| Open Datasets | Yes | We first evaluate the difference between explicit networks and (Implicit)2 networks on representing a high-resolution 512 512 grayscale image, which is a commonly used goalpost for evaluating implicit representation models from the scikit-image package [30]. ... We train the network on only 25% of the pixels from each image in the Natural and Text dataset, following [11]... Following [27], we train the models to fit a 7-second music piece. ... We choose three formulations of Fourier-MFN... and fitted them on several 3D object meshes with the point occupancy prediction objective similar to [18]. |
| Dataset Splits | No | The paper specifies training on 25% and evaluating on an unobserved 25% for image generalization, but does not explicitly mention a separate validation split or its details. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts) used for running its experiments. |
| Software Dependencies | No | The paper mentions software like 'scikit-image package [30]' but does not provide specific version numbers for any software components. |
| Experiment Setup | Yes | We train each model for 5000 iterations (under the same setting) using all pixels in the image (i.e., batch size 262,144)... we found that performing one step of the fixed-point iteration is sufficient... we propose to use a truncated backward gradient such that the Neumann series is unrolled for a T > 0 steps... In practice, we found T = 1 to be already sufficient... We additionally apply spectral normalization on the layers F we use for the implicit models. |