Benchmarking Deletion Metrics with the Principled Explanations

Authors: Yipei Wang, Xiaoqian Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we conduct experiments to 1) Validate TRACE s optimality and the capability of serving as the principled explanation; and 2) Use TRACE to assess the impact of different settings (as discussed in Section 3) to address the OOD concern in deletion metric.
Researcher Affiliation Academia 1Elmore Family School of Electrical and Computer Engineering, Purdue University, IN, USA. Correspondence to: Xiaoqian Wang <joywang@purdue.edu>.
Pseudocode Yes The associated pseudo-code is provided in Appendix D.
Open Source Code Yes The implementation of TRACE is available as open source at Git Hub.
Open Datasets Yes Experiments utilize the Image Net1k (ILSVRC2012) dataset (Deng et al., 2009) with images resized to 224 224.
Dataset Splits Yes The results are demonstrated in Figure 2(b). We present this comparison using TRACE-SA-Le Mo. And the resolution of both TRACE and the deletion metric is set to t = 7 7 = 49.
Hardware Specification Yes Experiments are carried out on Intel(R) Core(TM) i9-9960X CPU @ 3.10GHz with Quadro RTX 6000 GPUs.
Software Dependencies No The paper mentions "torchvision" but does not specify a version number or provide version numbers for any other software libraries or tools.
Experiment Setup Yes For SA, we use K = 5000 iterations, initial temperatures of T0 = 2 for -y and T0 = 0.1 for -p, and a cooling rate of η = 0.999.