Improving Non-Transferable Representation Learning by Harnessing Content and Style

Authors: Ziming Hong, Zhenyi Wang, Li Shen, Yu Yao, Zhuo Huang, Shiming Chen, Chuanwu Yang, Mingming Gong, Tongliang Liu

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, the proposed H-NTL significantly outperforms competing methods by a large margin.
Researcher Affiliation Collaboration Ziming Hong1 Zhenyi Wang2 Li Shen3 Yu Yao4,5 Zhuo Huang1 Shiming Chen5 Chuanwu Yang6 Mingming Gong5,7 Tongliang Liu1 1Sydney AI Centre, The University of Sydney 2University of Maryland, College Park 3JD Explore Academy 4Carnegie Mellon University 5Mohamed bin Zayed University of Artificial Intelligence 6Huazhong University of Science and Technology 7The University of Melbourne
Pseudocode Yes Algorithm 1 Train target-specified H-NTL
Open Source Code Yes For t NTL and s NTL, we use their released code7 and the same hyperparameters settings reported in their paper to run experiments. 7https://github.com/condition Wang/NTL
Open Datasets Yes The digit tasks contain three random-selected pairs from four digit datasets: MNIST (MT) (Deng, 2012), MNIST-M (MM) (Ganin et al., 2016), SVHN (SN) (Netzer et al., 2011) and SYN-D (SD) (Roy et al., 2018). For challenging tasks, we involve CIFAR10 to STL10 (Coates et al., 2011) (C10 S10), Vis DA (Peng et al., 2017) (VT VV) and Office Home (Venkateswara et al., 2017) (OP OC).
Dataset Splits Yes Following Wang et al. (2022b), we randomly select 8,000 samples as training data and 1,000 samples as testing data without overlap for digit tasks, C10 S10, and VT VT. For OP OC, we use 3,000 for training data and 1,000 samples due to the limitation of the dataset size.
Hardware Specification Yes Our code is implemented in Python 3.8.8 and Py Torch 1.8.0. All experiments are conducted on a server running Ubuntu 20.04 LTS, equipped with an NVIDIA RTX A6000 GPU.
Software Dependencies Yes Our code is implemented in Python 3.8.8 and Py Torch 1.8.0.
Experiment Setup Yes For training SL, we employ the SGD as an optimizer with lr = 0.001 and set the batch size to 32. ... For our proposed H-NTL, we employ the SGD as an optimizer with lr = 0.1 and set the batch size to 128. The disentanglement VAE is trained for 20 epochs, and the dual-path knowledge distillation is trained for 30 epochs. The hyper-parameter λt is set to 1.0 for all datasets.