This Looks Like That: Deep Learning for Interpretable Image Recognition

Authors: Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, Jonathan K. Su

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments show that Proto PNet can achieve comparable accuracy with its analogous non-interpretable counterpart, and when several Proto PNets are combined into a larger network, it can achieve an accuracy that is on par with some of the best-performing deep models. We demonstrate our method on the CUB-200-2011 dataset and the Stanford Cars dataset.
Researcher Affiliation Collaboration Chaofan Chen Duke University cfchen@cs.duke.edu Oscar Li Duke University oscarli@alumni.duke.edu Chaofan Tao Duke University chaofan.tao@duke.edu Alina Jade Barnett Duke University abarnett@cs.duke.edu Jonathan Su MIT Lincoln Laboratory su@ll.mit.edu Cynthia Rudin Duke University cynthia@cs.duke.edu
Pseudocode Yes The entire training algorithm is summarized in an algorithm chart, which can be found in Section S9.3 of the supplement.
Open Source Code Yes Supplementary Material and Code: The supplementary material and code are available at https: //github.com/cfchen-duke/Proto PNet.
Open Datasets Yes We trained and evaluated our network on the CUB-200-2011 dataset [45] of 200 bird species. We also trained our Proto PNet on the Stanford Cars dataset [20] of 196 car models.
Dataset Splits No The paper mentions "using cross validation" for choosing a hyperparameter (D) but does not provide specific details about train/validation/test splits, sample counts, or the methodology of the cross-validation itself.
Hardware Specification No The paper does not provide specific hardware details such as CPU/GPU models, memory, or cloud computing instances used for experiments.
Software Dependencies No The paper mentions using convolutional layers from models like VGG-16, VGG-19, ResNet-34, etc., but it does not provide specific version numbers for software libraries, frameworks, or operating systems used in the experiments.
Experiment Setup Yes For the bird dataset with input images resized to 224x224x3, the spatial dimension of the convolutional output is H = W = 7, and the number of output channels D in the additional convolutional layers is chosen from three possible values: 128, 256, 512, using cross validation. In our Proto PNet, we allocate a pre-determined number of prototypes mk for each class k (10 per class in our experiments).