Learning Human-Compatible Representations for Case-Based Decision Support

Authors: Han Liu, Yizhou Tian, Chacha Chen, Shi Feng, Yuxin Chen, Chenhao Tan

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

Reproducibility Variable Result LLM Response
Research Type Experimental Using both synthetic data and human subject experiments in multiple classification tasks, we demonstrate that such representation is better aligned with human perception than representation solely optimized for classification.
Researcher Affiliation Academia Han Liu, Yizhou Tian, Chacha Chen, Shi Feng, Yuxin Chen & Chenhao Tan Department of Computer Science, University of Chicago {hanliu,tianh,chacha,shif,chenyuxin,chenhao}@uchicago.edu
Pseudocode No The paper does not contain any pseudocode or algorithm blocks.
Open Source Code Yes Our code and data are available at https://github.com/Chicago HAI/ learning-human-compatible-representations.
Open Datasets Yes butterfly vs. moth classification from Image Net (Krizhevsky et al., 2012), and (ii) pneumonia classification based on chest X-rays (Kermany et al., 2018).
Dataset Splits Yes We generate 2000 images and randomly split the dataset into training, validation, and testing sets in a 60%:20%:20% ratio.
Hardware Specification Yes We use a computing cluster at our institution. We train our models on nodes with different GPUs including Nvidia Ge Force RTX2080Ti, Nvidia Ge Force RTX3090, Nvidia Quadro RTX 8000, and Nvidia A40.
Software Dependencies No We use the Py Torch framework (Paszke et al., 2019) and the Py Torch Lightning framework (Falcon et al., 2019) for implementation. (Specific version numbers are not provided.)
Experiment Setup Yes We use the Adam optimizer (Kingma & Ba, 2014) with learning rate 1e 4. We use a training batch size of 40 for triplet prediction, and 30 for classification.