Debugging and Explaining Metric Learning Approaches: An Influence Function Based Perspective

Authors: Ruofan Liu, Yun Lin, XIANGLIN YANG, Jin Song Dong

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In our quantitative experiment, EIF outperforms the traditional baseline in identifying more relevant training samples with statistical significance and 33.5% less time. In the field study on the well-known datasets such as CUB200, CARS196, and In Shop, EIF identifies 4.4%, 6.6%, and 17.7% labelling mistakes
Researcher Affiliation Academia Ruofan Liu Shanghai Jiao Tong University National University of Singapore Yun Lin Shanghai Jiao Tong University National University of Singapore Xianglin Yang National University of Singapore Jin Song Dong National University of Singapore
Pseudocode Yes Algorithm 1 Training Sample Relabelling
Open Source Code Yes Our code is available at https: //github.com/lindsey98/Influence_function_metric_learning.
Open Datasets Yes We use Proxy-NCA++ [31] and Soft Triple [23] loss to train DML models with Res Net-50 model architecture on three datasets, i.e., CUB200 [37], CARS196 [18], and In Shop [19].
Dataset Splits No The paper defines training and testing datasets explicitly (Xtrain and Xtest) but does not provide specific details on validation splits, percentages, or methodology for a separate validation set.
Hardware Specification No The paper does not provide specific details about the hardware used for running experiments, such as GPU models, CPU types, or memory specifications.
Software Dependencies No The paper mentions using specific models (Proxy-NCA++, Soft Triple, ResNet-50) but does not provide version numbers for programming languages, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow, CUDA) used in the implementation.
Experiment Setup No The paper mentions using specific models (Proxy-NCA++, Soft Triple, ResNet-50) and types of experiments, stating 'More details of training configuration can be referred on our website [2].' However, it does not explicitly provide concrete hyperparameter values or detailed training configurations within the main text.