논문리뷰: Repairing Deep Neural Networks: Fix Patterns and Challenges

(Loss function- add, remove or replace the loss function에 집중하여)

📕 Summary

Background

The paper presents a comprehensive study on bug fix patterns in deep neural networks (DNNs) used in software engineering, aiming to understand the challenges and patterns involved in repairing DNNs.

Problem (summarize)

The problem addressed in the research paper is the need for automated tools to support the repair of software that uses Deep Neural Networks (DNNs) .

Solution (This paper's keywork, key insight)

Summary of the bug fix patterns.

Loss function

The study on bug fix patterns in deep neural networks (DNNs) found that the loss function is a key parameter that is often modified during bug fixes in DNNs. This includes adding, removing, or updating the loss function during training. Different problems may require different loss functions, such as cross-entropy loss for classification problems and mean square error loss for regression-based problems. Custom loss functions may also be used for specific training results

MSE : (상대지표 ) 기본적으로 회귀 모형에서 많이 사용하는 Loss이자 성능 지표.
예측값-실제값의 차이에 대해, 평균 제곱합의 개념으로서, 낮을수록 좋은 성능 지표.

Algorithm(example)

Contribution

• The paper presents the results of a comprehensive study on bug fix patterns in deep neural networks (DNNs) used in software engineering .

• The study analyzed 415 repairs from Stack Overflow and 555 repairs from GitHub for popular deep learning libraries, including Caffe, Keras, Tensorflow, Theano, and Torch.

• The key findings reveal that DNN bug fix patterns are distinct from traditional bug fix patterns, with the most common patterns being fixing data dimension and neural network connectivity .