Improving Heterogeneous Model Reuse by Density Estimation

Authors: Anke Tang, Yong Luo, Han Hu, Fengxiang He, Kehua Su, Bo Du, Yixin Chen, Dacheng Tao

IJCAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on both synthetic and benchmark data demonstrate the superiority of the proposed method.
Researcher Affiliation Collaboration Anke Tang 1,2 , Yong Luo 1,2 , Han Hu 3 , Fengxiang He 4 , Kehua Su 1 , Bo Du 1,2 , Yixin Chen 5 , Dacheng Tao 6 1 School of Computer Science, National Engineering Research Center for Multimedia Software, Institute of Artificial Intelligence and Hubei Key Laboratory of Multimedia and Network Communication Engineering, Wuhan University, China 2 Hubei Luojia Laboratory, Wuhan, China 3 School of Information and Electronics, Beijing Institute of Technology, China 4 JD Explore Academy, JD.com, Inc., China 5 Department of CSE, Washington University in St. Louis, USA 6 The University of Sydney, Australia
Pseudocode Yes Algorithm 1 Heterogeneous Model Reuse aided by Density Estimation (without Calibration).
Open Source Code Yes The code is available at https://github.com/tanganke/HMR.
Open Datasets Yes We evaluate our method, HMR, and RKME on Fashion MNIST [Xiao et al., 2017], a popular benchmark dataset in the machine learning community, containing 70, 000 28 28 gray-scale fashion product images, each associated with a label from 10 classes.
Dataset Splits Yes The complete training set is split into a training set of 60, 000 examples and a test set of 10, 000 examples. To simulate the multiparty setting, we separate the training set into different parties with biased sample distribution.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU, GPU models, memory) used for running the experiments. It only mentions 'a 3-layer convolutional network' for local classifiers.
Software Dependencies No The paper mentions "scikit-learn package" but does not provide specific version numbers for it or any other software dependencies.
Experiment Setup Yes We set the training batch size to be 128, and the learning rate of all local models to 1e-4 during the local training. The learning rate is 1e-5 during the calibration step. All local classifiers have the same 3-layer convolutional network and all local density estimators are the same 12-layer real non-volume preserving (real NVP) flow network [Dinh et al., 2016]. As For RKME, we set the reduced dimension size to 10, and the number of generated samples to 200.