Post-hoc Part-Prototype Networks
Authors: Andong Tan, Fengtao Zhou, Hao Chen
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we first leverage commonly used explainability axioms (desired characteristics) in prior post-hoc methods mainly explaining where to show these axioms are easily fulfilled by our method, while prior prototype based models fail in sensitivity, completeness and linearity axioms qualitatively. We call a method fulfilling these axioms faithful explanations following (Wolf et al., 2023). Then we compare with existing part-prototype models in explaining what the model looks for and demonstrate the benefits quantitatively that our method yields more consistent and stable prototypes, yet guaranteeing the accuracy in the CUB-200-2011 benchmark dataset (Wah et al., 2011) in 5 backbones. In addition, we show our method can be easily applied to large scale datasets such as Image Net (Deng et al., 2009), while prior prototype based models either fail due to huge memory consumption or exhibit significant performance drops. Comprehensive ablation study is offered to discuss the importance and influence of our design choices. |
| Researcher Affiliation | Academia | 1Department of Computer Science and Engineering, Hong Kong University of Science and Technology, Hong Kong, China 2Center for Aging Science, HKUST, Hong Kong, China. |
| Pseudocode | No | The paper describes processes and steps, but does not include any formally structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper states it uses public code for some components (e.g., NMF from Collins et al., 2018), but it does not state that the code for its own proposed methodology is open-source or provide a link. |
| Open Datasets | Yes | Following prior works, we use the benchmark dataset CUB-200-2011 (Wah et al., 2011)... We leverage the Image Net (Deng et al., 2009) pretrained models... The Stanford Cars (Krause et al., 2013) and Stanford Dogs (Khosla et al., 2011) datasets are both for fine-grained image classification tasks. |
| Dataset Splits | No | The paper mentions using benchmark datasets but does not explicitly state the training, validation, or test split percentages or sample counts, nor does it cite a source defining specific splits for reproduction purposes. |
| Hardware Specification | No | The paper mentions a "NVIDIA 3090 Ti with 24 GB memory" in the context of other models' memory consumption, but it does not specify the hardware used for its own experiments. |
| Software Dependencies | No | For the prototype scaling via convex optimization, we use the cvxpy (Diamond & Boyd, 2016) package. We use the Sci Py (Virtanen et al., 2020) package for this optimization. The paper names software packages but does not provide specific version numbers directly in the text (e.g., "cvxpy 1.x.x" or "SciPy 1.0"). While the SciPy citation references "Sci Py 1.0", this is not explicitly stated in the body text when referring to the dependency. |
| Experiment Setup | Yes | We train 100 epochs with initial learning rate 0.001 and multiply it by 0.1 every 30 epochs. We use SGD as the optimizer with momentum 0.9 and weight decay 10 4. We only use horizontal flip with probability 0.5 as data augmentation. |