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..
SDPGO: Efficient Self-Distillation Training Meets Proximal Gradient Optimization
Authors: Tongtong Su, Yun Liao, Fengbo Zheng
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments on image classification, object detection, and semantic segmentation demonstrate that our method consistently surpasses recent state-of-the-art knowledge distillation techniques. ... In this paper, we conduct experiments on five visual recognition datasets, namely CIFAR-10/100 [24], Image Net [6], CUB200-2011 [44], and Cars196 [23]. We employ eight representative architectures [40] for evaluation... |
| Researcher Affiliation | Academia | Tongtong Su1, Liao Yun2 , Fengbo Zheng1 1School of Computer and Information Engineering, Tianjin Normal University 2College of Artificial Intelligence, Tianjin University of Science and Technology Corresponding author: Liao Yun (EMAIL), Fengbo Zheng (EMAIL) |
| Pseudocode | No | The paper includes mathematical equations (Eq. 1-10) and figures illustrating the framework (Figure 1, Figure 2) but does not contain explicitly labeled pseudocode or algorithm blocks describing the method steps in a structured, code-like format. |
| Open Source Code | Yes | Code is available at: https://github.com/nanxiaotong/SDGPO. |
| Open Datasets | Yes | In this paper, we conduct experiments on five visual recognition datasets, namely CIFAR-10/100 [24], Image Net [6], CUB200-2011 [44], and Cars196 [23]. ... To assess its generalization capability, we transfer our model to object detection task on the COCO-2017 dataset [29]. ... In addition, we employ the Pascal VOC, ADE20K, and Cityscapes benchmarks for semantic segmentation. |
| Dataset Splits | Yes | For Image Net, we use 1.2 million images for training and 50,000 images for validation. The MS-COCO dataset [29] has been established as a large-scale benchmark... Its curated collection consists of 118,000 images for model training and 5,000 images for validation. The ADE20K dataset [64] serves as a benchmark... comprising 20,210 training, 2,000 validation, and 3,352 testing images. The Cityscapes dataset [5]... comprises 5,000 high-resolution images... divided into 2,975 training, 500 validation, and 1,525 testing images. |
| Hardware Specification | Yes | We measure the actual speedup of lightweight neural networks on two mobile devices, Raspberry Pi 4B and Raspberry Pi 5 in Figure 4. ... All experiments are conducted on 8 NVIDIA Tesla-A100 GPUs. |
| Software Dependencies | No | The paper mentions using 'SGD optimizer' and general frameworks like 'deep neural networks' and 'computer vision tasks' but does not specify any particular software libraries, frameworks, or their version numbers (e.g., PyTorch 1.x, TensorFlow 2.x). |
| Experiment Setup | Yes | We used the Stochastic Gradient Descent (SGD) optimizer with a momentum of 0.9, a weight decay of 5e-4, and a temperature τ of 3. On CIFAR-100, we used a batch size of 64, an initial learning rate of 0.05, and trained the model for 240 epochs. For Image Net, training was conducted over 100 epochs, with learning rate adjustments at epochs 30, 60, and 90 by multiplying the learning rate by 0.1. For CUB200, the learning rate was set to 0.05 with a warmup period of 2 epochs, a batch size of 64, and a total of 240 training epochs. The learning rate was adjusted at epochs 70, 140, and 210 by multiplying it by 0.1, while all other hyperparameters remained constant. For Cars196, training was conducted for 200 epochs, with learning rate adjustments at epochs 60, 120, and 180 by multiplying it by 0.1. |