Defining and extracting generalizable interaction primitives from DNNs
Authors: Lu Chen, Siyu Lou, Benhao Huang, Quanshi Zhang
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that the extracted interactions can better reflect common knowledge shared by different DNNs1. ... We conducted experiments on various dataset. Experiments showed that our proposed method significantly improved the generalization power of the extracted interactions across different DNNs. |
| Researcher Affiliation | Academia | Lu Chen1 Siyu Lou1,2 Benhao Huang1 Quanshi Zhang1 1Shanghai Jiao Tong University, Shanghai, China 2Eastern Institute of Technology, Ningbo, China |
| Pseudocode | No | The paper describes its method in text and mathematical formulations but does not include a clearly labeled pseudocode or algorithm block. |
| Open Source Code | Yes | 1https://github.com/sjtu-xai-lab/generalizable-interaction |
| Open Datasets | Yes | We finetuned the pre-trained BERTBASE model and the BERTLARGE model on the SST-2 dataset (Socher et al., 2013) for sentiment classification. ... We explained the DNNs outputs on the SQu AD dataset (Rajpurkar et al., 2016). ... We extracted two sets of AND-OR interaction primitives from the Res Net-20 model (He et al., 2016) and the VGG-16 model (Simonyan & Zisserman, 2015), which were trained on the MNIST dataset (Le Cun, 1998). |
| Dataset Splits | No | The paper uses datasets like SST-2, SQuAD, and MNIST and reports classification accuracy, but it does not provide explicit training, validation, and test split percentages, sample counts, or refer to specific predefined splits of these datasets. |
| Hardware Specification | No | The paper reports average run-time per sample in Table 1, but it does not specify any hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. |
| Software Dependencies | No | The paper mentions various models like BERT, LLaMA, OPT, ResNet, and VGG, but it does not provide specific software dependencies or library versions (e.g., Python, PyTorch, TensorFlow versions) used in the experiments. |
| Experiment Setup | Yes | For each DNN, we further set a threshold τ (i) to collect a set of salient interactions from this DNN as interaction primitives, i.e., τ (i) = 0.05 max S |I(S|x)|. ... We took the most salient k interactions from each i-th DNN as the set of AND-OR interaction primitives, i.e., |Ωand, (i)| = |Ωor, (i)| = k... For comparison, we set the value of α to [0, 0.2, 0.4, 0.6, 0.8, 1.0], respectively. |