Gaussian Mixture Convolution Networks

Authors: Adam Celarek, Pedro Hermosilla, Bernhard Kerbl, Timo Ropinski, Michael Wimmer

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental A thorough evaluation of GMCNs, including ablation studies. and To evaluate our architecture, we use the proposed GMCN architecture to train classification networks on a series of well-known tasks.
Researcher Affiliation Academia TU Wien Ulm University
Pseudocode No The paper describes methods and equations, but does not include explicit pseudocode blocks or algorithms labeled as such.
Open Source Code Yes The source code for a proof-of-concept implementation, instructions, and benchmark datasets are provided in our Git Hub repository (https://github.com/cg-tuwien/ Gaussian-Mixture-Convolution-Networks).
Open Datasets Yes The MNIST data set (Lecun et al., 1998) and Model Net10 (Zhirong Wu et al., 2015)
Dataset Splits Yes We trained our GMCN architecture using the standard train/test splits by fitting 64 Gaussians to each image and processing them with our model. and We trained our network on Model Net10 using the standard train/test splits without data augmentation.
Hardware Specification Yes We used an NVIDIA Ge Force 2080Ti for training.
Software Dependencies No The paper mentions software like Adam optimizer, PyTorch, Tensor Board, and autodiff C++ library (Leal, 2018), but does not provide specific version numbers for these dependencies.
Experiment Setup Yes All models are trained using the Adam optimizer, with an initial learning rate of 0.001. The learning rate is reduced by a scheduler once the accuracy plateaus. Moreover, we apply weight decay scaled by 0.1 of the learning rate to avoid overfitting, as outlined in Section 4.2.