comparison

Prepare for Reviewer B

An intuitive comparison of DeepArc with other approaches is as follows.

Prepare for Reviewer C

An example of a regression task 

We take a ResNet101  trained on VOC2007  to solve object detection as an example. 

The model consists of  33 blocks and 101 convolution layers. It can be modularized to {· · · , [6−7], · · · , [19−30], · · · ,[32 − 33]}  . 

There are about 36% of the layers can be encapsulated for benefit .

Why choose this example ?

In fact, there are many classic regression tasks.  Here are some examples.

1.  Given a set of unordered features that describe one thing, we need to predict a continuous real value. For example, Boston house price prediction.

These are usually solved with classical machine learning models such as Linear Regression, Random Forrest Regressor, and XGBoost Regressor (NOT CNN).

2.  Given a set of ordered features such as a temporal sequence of one thing, we need to predict the sequence in the next time step. For example, Stock Market Prediction.

These are usually solved with recurrent neural networks such as simple RNN, LSTM and GRU (NOT CNN).

3.  Some other emerging tasks such as object detection include classification(predicting category) and regression(predicting coordinates) tasks. (Can be solved by CNN)

So here we take a ResNet101  trained on VOC2007  to solve object detection as an example. 

Why is it not presented in the paper?

The models we analyzed are some classic CNNs(Convolutional Neural Networks) such as VGG, ResNet and Wide-ResNet, and these focus on classification tasks.

In fact, in the field of SE for AI, most papers use DNN/CNN for classification tasks to prove the effectiveness of their techniques. For example, two modular works ESEC/FSE20(distinguished paper) and ICSE22 , some model reuse works such as ISSTA21 and ICSE22 , DNN testing works such as ISSTA19 and  TOSEM21

Thanks again for these helpful insights to improve our work!