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 [1].

SelKD: Selective Knowledge Distillation via Optimal Transport Perspective

Authors: Liangliang Shi, Zhengyan Shi, Junchi Yan

ICLR 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on standard benchmarks demonstrate the superiority of our approach. The source code is available at: https://github.com/machoshi/Sel KD. [...] Our experiments are performed using Py Torch 1.4.0 and run on Intel Core i7-7820X CPU @ 3.60GHz with Nvidia Ge Force RTX 3080. We take single GPU for classification on CIFAR-10, CIFAR-100 (Krizhevsky et al., 2009) and Tiny Image Net (Le & Yang, 2015), and evaluate on testing data by top-1 accuracy. [...] The results of Sel KD tasks on the CIFAR-100 and Tiny Image Net dataset are presented in Table 1 and Table 2, with results for CIFAR-10 is shown in the Appendix A. [...] Table 4: Ablation Study of Sel KD on CIFAR-100.
Researcher Affiliation Academia 1Sch. of Computer Science & Sch. of Artificial Intelligence, Shanghai Jiao Tong University 2Shanghai Artificial Intelligence Laboratory Equal Contributions Correspondence Author EMAIL
Pseudocode Yes Algorithm 1: Computing the KL divergence between student and teacher couplings under Open-set Sel KD
Open Source Code Yes Experimental results on standard benchmarks demonstrate the superiority of our approach. The source code is available at: https://github.com/machoshi/Sel KD.
Open Datasets Yes Our experiments are performed using Py Torch 1.4.0 and run on Intel Core i7-7820X CPU @ 3.60GHz with Nvidia Ge Force RTX 3080. We take single GPU for classification on CIFAR-10, CIFAR-100 (Krizhevsky et al., 2009) and Tiny Image Net (Le & Yang, 2015), and evaluate on testing data by top-1 accuracy.
Dataset Splits No The paper describes how classes are divided into subtasks (e.g., "CIFAR-100 and Tiny Image Net, with 100 classes each, are divided into 5 subtasks, each containing 20 classes"), but it does not specify the train/test/validation dataset splits (e.g., 80/10/10 percentages or sample counts) for these datasets. It only mentions evaluating on "testing data" without detailing the splitting methodology for the main datasets.
Hardware Specification Yes Our experiments are performed using Py Torch 1.4.0 and run on Intel Core i7-7820X CPU @ 3.60GHz with Nvidia Ge Force RTX 3080.
Software Dependencies Yes Our experiments are performed using Py Torch 1.4.0 and run on Intel Core i7-7820X CPU @ 3.60GHz with Nvidia Ge Force RTX 3080.
Experiment Setup Yes As for the learning rate, we set 0.05 for all tasks with regard to CIFAR-10 and CIFAR-100 datasets. For Tiny Image Net, learning rate is 0.2.