Learning Single Index Models with Diffusion Priors
Authors: Anqi Tang, Youming Chen, Shuchen Xue, Zhaoqiang Liu
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform numerical experiments on image datasets for different nonlinear measurement models. We observe that compared to competing methods, our approach can yield more accurate reconstructions while utilizing significantly fewer neural function evaluations. In this section, we conduct a series of experiments to validate the effectiveness of the proposed SIM-DMIS approach (see Algorithm 1). Specifically, we evaluate our method on two datasets: FFHQ 256 256 (Karras et al., 2019), Image Net 256 256 (Deng et al., 2009). |
| Researcher Affiliation | Academia | 1University of Electronic Science and Technology of China 2University of Chinese Academy of Sciences 3Academy of Mathematics and Systems Science, CAS. Correspondence to: Zhaoqiang Liu <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 The SIM-DMIS approach Input: A Rm n, y Rm, xθ, time steps ϵ = t N < t N 1 < . . . < t1 < t0 = T, generator G corresponding to the sampling process as in Eq. (14), partial inversion operator G t as in Eq. (19), tuning parameters Cs and C s 1: Calculate t [ϵ, T] as in Eq. (26). 2: Calculate the estimated vector ˆx as in Eq. (27). Output: ˆx |
| Open Source Code | No | The paper does not contain an explicit statement about the release of its source code, nor does it provide a link to a code repository. Mentions of code refer to third-party pre-trained models used in the experiments. |
| Open Datasets | Yes | Specifically, we evaluate our method on two datasets: FFHQ 256 256 (Karras et al., 2019), Image Net 256 256 (Deng et al., 2009). For FFHQ and Im-age Net, the ambient dimension is n = 3 256 256 = 196608. the results of CIFAR-10 (Krizhevsky & Hinton, 2009) are included in Appendix D. |
| Dataset Splits | Yes | The reported quantitative results are averaged over 100 testing images. In the experimental process, we use DDIM for sampling and the DM2M inversion method (refer to Sec. 2.1 for details). ... We employ the model from DPS, which has been trained on 49,000 FFHQ 256 256 images and validated on the first 1,000 images. ... The tests are conducted on 100 images selected from the Validation1K Set. ... For 1-bit measurements, we report results under two measurement settings: m = 500 and m = 1000, corresponding to approximately 16% and 32% of the ambient dimension n. |
| Hardware Specification | Yes | All experiments are executed on a single NVIDIA Ge Force RTX 4090 GPU. |
| Software Dependencies | No | The paper mentions software such as PyTorch, and references various models (DPS, ADM, DDPM) and methods (DDIM, DM2M), but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | For SIM-DMS, we use 50 NFEs, while for SIM-DMIS and SIM-DMFIS, we use 150 NFEs. The remaining parameters are set as ϵ = 0.001, T = 1, and σ = 0.05. For the detailed parameter settings of Cs and C s in SIM-DMS and SIM-DMIS, please refer to Appendices F and G. Table 7. Ablation experiments for SIM-DMS on CIFAR-10 with m = 500 1-bit measurements, with different values of Cs, C s and NFEs. |