Reducing Flipping Errors in Deep Neural Networks
Authors: Xiang Deng, Yun Xiao, Bo Long, Zhongfei Zhang6506-6514
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on different benchmark datasets with different modern network architectures demonstrate that the proposed flipping error reduction (FER) approach can substantially improve the generalization, the robustness, and the transferability of DNNs without introducing any additional network parameters or inference cost, only with a negligible training overhead. |
| Researcher Affiliation | Collaboration | Xiang Deng1*, Yun Xiao2, Bo Long2, Zhongfei Zhang1 1 Computer Science Department, State University of New York at Binghamton 2 JD.com xdeng7@binghamton.edu, xiaoyun1@jd.com, bo.long@jd.com, zhongfei@cs.binghamton.edu |
| Pseudocode | No | The paper describes the framework and calculations using equations and text, but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | 1Code: https://github.com/Xiang-Deng-DL/FER |
| Open Datasets | Yes | CIFAR-100 (Krizhevsky and Hinton 2009) and Standard Dogs (Khosla et al. 2011). Tiny Image Net 2https://tiny-imagenet.herokuapp.com CUB-200-2011 (Wah et al. 2011), Stanford Dogs (Khosla et al. 2011), and FGVC-Aircraft (Maji et al. 2013), two tabular datasets, i.e., Arcene (Dua and Graff 2017), and Iris (Dua and Graff 2017). |
| Dataset Splits | No | A DNN is typically trained for many epochs and then a validation dataset is used to select the DNN in an epoch (we simply call this epoch the last epoch ) as the final model for making predictions on unseen samples, while it usually cannot achieve a perfect accuracy on unseen samples. The paper uses standard benchmark datasets which often have predefined splits, but it does not explicitly state the specific sizes or percentages of these splits within the paper itself. |
| Hardware Specification | No | The paper discusses the training of DNNs and various experiments but does not provide specific details on the hardware, such as GPU or CPU models, used for computation. |
| Software Dependencies | No | The paper mentions using open-source implementations for certain tasks (e.g., 'open-source implementation 3 of (Zhang et al. 2017)') but does not specify version numbers for any software dependencies or libraries. |
| Experiment Setup | Yes | It is observed that τ = 5 consistently outperforms τ = 10. The reason is that as seen from Eq. (2), a large τ makes the behavior too flat to contain sample-to-class distance information. On the other hand, we observe that the performance of FER is not sensitive to µ. In most cases, we simply set α and β to 1.0 − 0.9 k M and 0.9 k M (where M is the number of total training epochs), respectively, based on the fact that the average behavior becomes more accurate in the later epoch with the DNN learning more information. |