Sparse Bayesian Generative Modeling for Compressive Sensing

Authors: Benedikt Böck, Sadaf Syed, Wolfgang Utschick

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

Reproducibility Variable Result LLM Response
Research Type Experimental We support our approach theoretically through the concept of variational inference and validate it empirically using different types of compressible signals. and 4 Experiments
Researcher Affiliation Academia Benedikt Böck, Sadaf Syed, Wolfgang Utschick TUM School of Computation, Information and Technology Technical University of Munich {benedikt.boeck,sadaf.syed,utschick}@tum.de
Pseudocode Yes N Pseudo-Code for the Training and Inference of the CSVAE and CSGMM
Open Source Code Yes 1Source code is available at https://github.com/beneboeck/sparse-bayesian-gen-mod.
Open Datasets Yes We use the MNIST dataset (N = 784) for evaluation [42, (CC-BY-SA 3.0 license)]. and We also use a dataset of 64 × 64 cropped celeb A images (N = 3 × 642 = 12288) [44] and evaluate on the Fashion MNIST dataset (N = 784) in Appendix L [45, (MIT license)].
Dataset Splits Yes We once reduce the learning rate by a factor of 2 during training and stop the training, when the modified ELBO in (15) for a validation set of 5000 samples does not increase.
Hardware Specification Yes All models have been simulated on an NVIDIA A40 GPU except for the proposed CSGMM, whose experiments have been conducted on an Intel(R) Xeon(R) Gold 6134 CPU @ 3.20GHz.
Software Dependencies No The paper mentions software like 'Adam' for optimization and 'Pywavelets' for wavelet analysis, but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes For CSGMM, we set the number K of components to 32... The CSVAE encoders and decoders contain two fully-connected layers with Re LU activation... The latent dimension is set to 16, the learning rate is set to 2 × 10−5, and the batch size is set to 64. We use the Adam optimizer for optimization [47].