QCS-SGM+: Improved Quantized Compressed Sensing with Score-Based Generative Models

Authors: Xiangming Meng, Yoshiyuki Kabashima

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments are conducted, demonstrating the substantial superiority of QCS-SGM+ over QCS-SGM for general sensing matrices beyond mere row-orthogonality.
Researcher Affiliation Academia 1The Zhejiang University-University of Illinois Urbana-Champaign Institute, Zhejiang University, China 2Institute for Physics of Intelligence & Department of Physics, The University of Tokyo, Japan xiangmingmeng@intl.zju.edu.cn, kaba@phys.s.u-tokyo.ac.jp
Pseudocode Yes Algorithm 1: QCS-SGM+ Input: {βt}T t=1, ϵ, γ, Iter EP, K, y, A, σ2, quantization thresholds {[lq, uq)|q Q} Initialization: x0 1 U (0, 1) 1 for t = 1 to T do 2 αt ϵβ2 t /β2 T 3 for k = 1 to K do 4 Draw zk t N(0, I) Initialization: h F , τ F , h G, τ G 5 for it = 1 to Iter EP do 7 τ G = 1 χa τ F 9 τ F = 1 χb τ G 10 Compute xt log p(y | xt) as (23) 11 xk t = xk 1 t + αt h sθ(xk 1 t , βt) + γ xt log p(y | xt) i + 2αtzk t 12 x0 t+1 x K t Output: ˆx = x K T
Open Source Code Yes The source code is available at https://github.com/mengxiangming/QCS-SGM-plus.
Open Datasets Yes Datasets: We consider three popular datasets: MNIST (Le Cun and Cortes 2010) , CIFAR-10 (Krizhevsky and Hinton 2009), and Celeb A (Liu et al. 2015).
Dataset Splits No The paper uses standard datasets but does not explicitly state the train/validation/test splits or their methodology for splitting.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used to run the experiments.
Software Dependencies No The paper mentions using NCSNv2 but does not provide specific version numbers for it or any other software dependencies.
Experiment Setup Yes The paper specifies various experimental settings and hyperparameters, including the number of measurements (M), noise standard deviation (σ), condition number (κ), and algorithm parameters such as {βt}T t=1, ϵ, γ, Iter EP, K, and quantization thresholds {[lq, uq)|q Q} as shown in Algorithm 1 and figure captions (e.g., Figure 2: 'M = 400, σ = 0.05, κ = 103' for MNIST).