Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Efficient and Generalizable Mixed-Precision Quantization via Topological Entropy
Authors: Nan Li, Yonghui Su, Lianbo Ma
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that the quantization policy obtained on CIFAR-10 can generalize to Image Net and PASCAL VOC. GMPQ-TE achieves a competitive accuracy-complexity trade-off compared to state-of-the-art MPQ methods. We conduct extensive experimental results on image classification and object detection, which show that the quantization policy generalized from CIFAR-10 to Image Net and PASCAL VOC achieves a competitive accuracy-complexity trade-off compared with the state-of-the-art MPQ methods. |
| Researcher Affiliation | Academia | Nan Li School of Artificial Intelligence, Institute of Big Data Science and Industry Shanxi University College of Software, Northeastern University EMAIL. Yonghui Su College of Software, Northeastern University EMAIL. Lianbo Ma College of Software, Northeastern University EMAIL. All listed institutions (Shanxi University, Northeastern University) are academic institutions. |
| Pseudocode | No | The paper describes the methodology in Section 3 'Methodology' using text and mathematical formulations. Figure 3 provides an overall framework illustration, but there are no explicitly labeled pseudocode or algorithm blocks with structured steps. |
| Open Source Code | No | The paper makes statements like 'is implemented by ourselves using open source code' or 'is implemented based on the quantized architectural information provided by the authors due to lack of open source code' in Section 4.3, but these refer to other methods being compared against, not the authors' own GMPQ-TE code. There is no explicit statement or link confirming the release of source code for GMPQ-TE. |
| Open Datasets | Yes | Following studies Wang et al. [2021] and Wang et al. [2024] configuration, for image classification, we use Image Net Deng et al. [2009] to evaluate the quantized networks... In terms of object detection, PASCAL VOC Everingham et al. [2015] is employed... We use CIFAR-10 to find the suitable quantization policy for all baseline architectures. |
| Dataset Splits | No | The paper mentions using CIFAR-10, Image Net, and PASCAL VOC, and refers to fine-tuning on 'different tasks' and obtaining quantization policies on CIFAR-10 for deployment on Image Net or PASCAL VOC. It also mentions selecting 'a batch of images with the label cat' for ablation studies and discusses 'different batch sizes'. However, specific numerical details for training, validation, and test splits (e.g., percentages or exact counts) for any of the datasets are not explicitly provided in the main text. |
| Hardware Specification | Yes | All experiments are perform on an NVIDIA Geforce GTX 3090Ti. Table 3 compares the hardware performance of different quantization methods on two FPGA platforms (XC7Z020 and XC7Z045)... On RTX 3090, Mobile Net-V2 runs at over 800 FPS with only 1.2 ms latency. Even on Jetson Nano, it maintains real-time performance under a 10W power budget. |
| Software Dependencies | No | The paper does not explicitly state specific version numbers for any software dependencies (e.g., programming languages, deep learning frameworks, or libraries). |
| Experiment Setup | Yes | We set different batch sizes (i.e., 32, 64, 128, 256, and 512) on CIFAR-10 to verify the effect of batch size on quantification sensitivity, where Res Net-18 as a baseline model is used to test the TE of each layer. [...] we consider the computational efficiency and choose 128 in all experiments. By setting different constraints T , we can obtain the quantized networks with different accuracy-complexity trade-offs. In addition, we follow study Wang et al. [2021] to fine-tune the quantized networks that are found on different tasks. |