ICACT20240217 Question.1
Questioner: jjhu@nttu.edu.tw 2024-02-05 오전 8:46:43 |
ICACT20240217 Answer.1
Answer by Auhor ristar1234@skku.edu 2024-02-05 오전 8:46:43
Chrome  Click!! |
| The paper proposes an improved method - Z-score based Neuron Coverage (ZSNC). This method uses z-score approximation to estimate the range of neuron outputs for each class, instead of extracting exact percentile values.
Experimental results show that the proposed ZSNC outperforms other NC derivative methods in multiple image classification benchmarks, both in effectiveness and efficiency.
Could using z-scores lead to some loss of information? |
Thank you for your kind summary of our ZSNC.
Since ZSNC uses approximation values rather than exact percentile values, it cannot be said that there is no information loss at all. Therefore, from the results of experiments A and G, we can see that the performance of the proposed method is slightly lower than that of previous studies.
However, it is also not guaranteed that finding the exact percentile value based on the training data is more advantageous in finding false predictions. One evidence is that the range performed through approximation showed higher effectiveness in all experiments except experiments A and G. Therefore, the proposal ZSNC is a method that could show high performance without high cost. |
ICACT20240217 Question.2
Questioner: jjhu@nttu.edu.tw 2024-02-05 오전 8:49:34 |
ICACT20240217 Answer.2
Answer by Auhor ristar1234@skku.edu 2024-02-05 오전 8:49:34
Chrome Click!! |
| Besides computational efficiency, are there any other potential advantages to using z-score approximation? |
In addition to efficiency, which measures the operating speed of the algorithm, ZSNC showed an average of more than 15% higher performance in 8 experiments than other NC-based methods in false prediction detection performance. Therefore, ZSNC is more advantageous in selecting inputs that may be weak samples for the learning model. |
ICACT20240217 Question.3
Questioner: jjhu@nttu.edu.tw 2024-02-05 오전 8:57:05 |
ICACT20240217 Answer.3
Answer by Auhor ristar1234@skku.edu 2024-02-05 오전 8:57:05
Chrome Click!! |
| Are there any plans to further optimize the inference efficiency of ZSNC in the future? |
To increase efficiency, we eliminated unnecessary steps during the operation of FD+, the latest NC-based research, and simplified pattern extraction through approximation.
Based on ResNet50, which is currently widely used as a backbone, it takes about 7 seconds to process 10,000 images. This means 1400 FPS, so future work to improve efficiency has not yet been considered. However, it seems necessary to conduct one more experiment at the ImageNet scale to confirm the effectiveness of real-world image resolution. Thank you for the great comment. |
ICACT20240217 Question.4
Questioner: madbrogada@bicol-u.edu.ph 2024-02-06 오후 2:06:08 |
ICACT20240217 Answer.4
Answer by Auhor ristar1234@skku.edu 2024-02-06 오후 2:06:08
Chrome Click!! |
| Are there any potential limitations or areas for further improvement identified during the evaluation of the proposed method? |
|
ICACT20240217 Question.5
Questioner: jllovido@bicol-u.edu.ph 2024-02-07 오후 12:02:07 |
ICACT20240217 Answer.5
Answer by Auhor ristar1234@skku.edu 2024-02-07 오후 12:02:07
Chrome Click!! |
| How generalizable is the proposed technique and to what extent can such technique be adapted or customized for different types of DNN classifiers beyond the datasets considered in the study? |
|
