Triple Wins: Boosting Accuracy, Robustness and Efficiency Together by Enabling Input-Adaptive Inference
Authors: Ting-Kuei Hu, Tianlong Chen, Haotao Wang, Zhangyang Wang
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 EXPERIMENTAL RESULTS, We evaluate accuracy, robustness, and efficiency, using the metrics below: Testing Accuracy (TA):... Adversarial Testing Accuracy (ATA):... Mega Flops (MFlops):..., Table 1: Benchmarking results of adverserial training of three networks., Table 2: The performance of RDI-Small CNN., Table 3: The performance evaluation on RDI-Res Net38., Table 4: The performance evaluation on RDI-Mobilenet V2. |
| Researcher Affiliation | Academia | Department of Computer Science and Engineering Texas A&M University, USA {tkhu,wiwjp619,htwang,atlaswang}@tamu.edu |
| Pseudocode | No | No pseudocode or algorithm blocks were found in the paper. Methods are described through mathematical formulas and textual explanations. |
| Open Source Code | Yes | The codes can be referenced from https://github.com/TAMU-VITA/triple-wins. |
| Open Datasets | Yes | We evaluate three representative CNN models on two popular datasets: Small CNN on MNIST (Chen et al., 2018); Res Net-38 (He et al., 2016) and Mobile Net-V2 (Sandler et al., 2018) on CIFAR-10. |
| Dataset Splits | Yes | To choose {ti}K i=1, Huang et al. (2018) provides a good starting point by fixing exiting probability of each branch classifiers equally on validation set so that each sample can equally contribute to inference. |
| Hardware Specification | No | No specific hardware details (e.g., GPU models, CPU types, memory) used for running experiments were mentioned in the paper. The paper discusses 'resource-constrained applications' and 'IoT devices' in a general context, but not for its own experimental setup. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies or libraries (e.g., PyTorch, TensorFlow, CUDA, or specific Python libraries) used in their experiments. |
| Experiment Setup | Yes | A LEARNING DETAILS OF RDI-NETS, We train for 13100 iterations with a batch size of 256. The learning rate is initialized as 0.033 and is lowered by 10 at 12000th and 12900th iteration. For hybrid loss, the weights {wi}N+1 i=1 are set as {1, 1, 1} for simplicity. For adversarial defense/attack, we perform 40-steps PGD for both defense and evaluation. The perturbation size and step size are set as 0.3 and 0.01. For RDI-Res Net38, we initialize learning rate as 0.1 and decay it by a factor of 10 at 32000th and 48000th iteration. The learning procedure stops at 55000 iteration. For RDI-Mobile Net V2, the learning rate is set to 0.05 and is lowered by 10 times at 62000th and 70000th iteration. We stop the learning procedure at 76000 iteration. For hybrid loss, we follow the discussion in (Hu et al., 2019) and set {wi}N+1 i=1 of RDI-Res Net38 and RDI-Mobile Net V2 as {0.5, 0.5, 0.7, 0.7, 0.9, 0.9, 2} and {0.5, 0.5, 1}, respectively. For adversarial defense/attack, the perturbation size and step size are set as 8/255 and 2/255. 10-steps PGD is performed for defense and 20-steps PGD is utilized for evaluation. |