Efficiency for Free: Ideal Data Are Transportable Representations

Authors: PENG SUN, Yi Jiang, Tao Lin

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments across four widely-used datasets, seven neural network architectures, eight self-supervised learning algorithms demonstrate the effectiveness and efficiency of RELA.
Researcher Affiliation Academia Peng Sun1,2 Yi Jiang1 Tao Lin2, 1Zhejiang University 2Westlake University sunpeng@westlake.edu.cn, yi_jiang@zju.edu.cn, lintao@westlake.edu.cn
Pseudocode Yes Algorithm 1 Adaptive Loss Weighting Algorithm for RELA and Self-Supervised Learning
Open Source Code Yes Our code is available at: https://github.com/LINs-lab/Re LA.
Open Datasets Yes Datasets: For low-resolution data (32x32), we evaluate our method on two datasets, i.e., CIFAR-10 [35] and CIFAR-100 [34]. For high-resolution data, we conduct experiments on two large-scale datasets including Tiny-Image Net (64x64) [36] and full Image Net-1K (224x224) [20]
Dataset Splits No The paper refers to 'Validation Loss' in figures, implying a validation set is used, but does not explicitly provide specific dataset split information (exact percentages, sample counts, or detailed splitting methodology) for training, validation, or testing.
Hardware Specification Yes We implement our method through Py Torch [48], and all experiments are conducted on NVIDIA RTX 4090 GPUs.
Software Dependencies No The paper mentions 'PyTorch [48]' but does not provide a specific version number for this or any other software library.
Experiment Setup Yes This includes using Adam W as the optimizer, with a mini-batch size of 128 (except for Image Net-1K, where we use a mini-batch size of 512). ... Optimizer Adam W, Learning Rate 0.001, Weight Decay 0.01. ... Optimizer Adam (learning rate: 3e-4)