Machine Learning Approach Regarding the Classification and Prediction of Dog Sounds: A Case Study of South Indian Breeds

Round 1

Reviewer 1 Report

Authors recorded and classified the barking patterns of Rajapalayam Hound and Kombai Hound, two popular south Indian breeds, in order to identify the barking pattern during different circumstances. The goal is to determine the context of the dog barking pattern based on several real-time events, such as whether the canines are alone, looking at strangers, or displaying a desire to fight. The subject is highly intriguing. I'd want to thank the authors for working on such an interesting topic. In any case, these are my thoughts:

1-Why are you employing the q-learning method? It should be noted in the abstract section.

2-The introduction part lacks the paper's structure. It must be mentioned.

3-It becomes necessary to discuss the importance of the DL methods in several fields. The authors are encouraged to use these crucial references when discussing it.

https://link.springer.com/article/10.1007/s00521-022-07424-w

https://www.sciencedirect.com/science/article/abs/pii/S2210670722004061

https://www.mdpi.com/1424-8220/22/4/1335

https://www.mdpi.com/2078-2489/13/4/203

4-If formulas are taken from other articles, they must be appropriately cited.

5-Several issues should be discussed in the conclusion section, including the limitations of the work, the disadvantages of the method, and the implications of the work outside the scope.

Author Response

Reviewer Question:1. Why are you employing the q-learning method? It should be noted in the abstract section.

Response:

We thank the reviewer for the compliment and comments. Q- Learning is reinforcement learning which will generate the next best action for the given state. It is model-free learning used to find the best course of dog action behaviour for the given current state of the dog. The relevant texts are added to the revised manuscript

Reviewer Question:2. The introduction part lacks the paper's structure. It must be mentioned.

Response:

We thank the reviewer for the comment. The overall structure of the paper was organized as follows, data collection and analysis were briefed in section 2. In section 3, the architecture and methodology of the dog sound classification with the learning algorithm were explained. The results of the proposed work were discussed in section 4. Section 5 concludes the work done. A text to discuss the structure is added in the revised manuscript

Reviewer Question:3. It becomes necessary to discuss the importance of the DL methods in several fields. The authors are encouraged to use these crucial references when discussing it.

https://link.springer.com/article/10.1007/s00521-022-07424-w

https://www.sciencedirect.com/science/article/abs/pii/S2210670722004061

https://www.mdpi.com/1424-8220/22/4/1335

https://www.mdpi.com/2078-2489/13/4/203

 

Response:

Discussed the above-mentioned journals as reference numbers 39, 40, 41, and 42 on page number 2 of the revised manuscript.

 

Reviewer Question:4.-If formulas are taken from other articles, they must be appropriately cited.

Response: Formulas cited.

 

Reviewer Question:5. Several issues should be discussed in the conclusion section, including the limitations of the work, the disadvantages of the method, and the implications of the work outside the scope.

Response:

Limitations of the work include the behaviour of the dogs in different scenarios such as different environments, climate factors, etc. need to be analysed. Different breeds of dogs can be taken for the study to identify the behavioural nature of the dogs. The text is also added to the revised manuscript

Reviewer 2 Report

1. The technical details of the paper need to be supplemented, including the explanation of some concepts and the introduction of the model structure, e.g. Q-learning method.

2. The convergence process of the model objective function needs to be shown to facilitate the reader’s understanding of the training situation.

3. The specific feature extraction process should be described carefully.

4. The hyper-parameters setting of the machine learning model is ignored.

5. The overall structure of the paper should be adjusted to be more clear.

6. The experimental part should give more detailed results, such as comparative ablation test under different parameter settings.

7. The key point of machine learning is whether it can be well generalized in the unknown environment. Therefore, the author should refer to and discuss the regularization and generalization of machine learning in the first section, such as “Cloud shape classification system based on multi-channel cnn and improved fdm.” IEEE Access 8 (2020): 44111-44124. “GPU-accelerated Faster Mean Shift with Euclidean distance metrics.” IEEE 46th Annual Computers, Software, and Applications Conference (COMPSAC). IEEE, 2022. “Deep Facial Diagnosis: Deep Transfer Learning From Face Recognition to Facial Diagnosis,” in IEEE Access, vol. 8, pp. 123649-123661, 2020. “Improvement of generalization ability of deep CNN via implicit regularization in two-stage training process,” IEEE Access, vol. 6, pp. 15844-15869, 2018.

Author Response

Reviewer Question:1. The technical details of the paper need to be supplemented, including the explanation of some concepts and the introduction of the model structure, e.g. Q-learning method.

Response:

We thank the reviewer for the comment. Q-learning algorithm is the descriptive form of reinforcement learning algorithms. Q-learning can learn an efficient technique even without an operating prototype by adjusting the reward and action of a state called Q function. It will find the next best action for the given current state which aims to choose the action at random and maximizes the reward. The relevant sections with the Q-learning algorithm are revised in the revised manuscript.

Reviewer Question:2. The convergence process of the model objective function needs to be shown to facilitate the reader’s understanding of the training situation.

Response:

We thank the reviewer for the comment. In the training process, the environment for the dog's behaviour had been set up such as the dog is placed inside the house with the owner. Then the observation of the agent(dog) is recorded to analyze its behaviour. Based on the instructions from the owner and environment dogs make action, exhibiting their behaviour. On observing the action of the dogs, rewards will be mapped to define a policy.

Reviewer Question:3. The specific feature extraction process should be described carefully.

Response: We thank the reviewer for the comment. MFCC is used in the feature extraction process and is discussed in the revised manuscript.

Reviewer Question:4. The hyper-parameters setting of the machine learning model is ignored.

Response:

We thank the reviewer for the comment. The process is repeated with the same data sets for the training and tests to compare different hyper-parameters. The hyperparameters include actor learning rate, exploration and environment.

Reviewer Question:5. The overall structure of the paper should be adjusted to be more clear.

Response: The structure of the paper is revised as recommended by the reviewer 1 and 2

Reviewer Question:6. The experimental part should give more detailed results, such as a comparative ablation test under different parameter settings.

Response: We thank the reviewer for the comment. Experimental results on different scenarios were discussed. Since this article was a behaviour description of two particular dogs, the models' performance cannot be compared with any approaches under different parameter settings.

Reviewer Question:7. The key point of machine learning is whether it can be well generalized in an unknown environment. Therefore, the author should refer to and discuss the regularization and generalization of machine learning in the first section, such as “Cloud shape classification system based on multi-channel cnn and improved fdm.” IEEE Access 8 (2020): 44111-44124. “GPU-accelerated Faster Mean Shift with Euclidean distance metrics.” IEEE 46th Annual Computers, Software, and Applications Conference (COMPSAC). IEEE, 2022. “Deep Facial Diagnosis: Deep Transfer Learning From Face Recognition to Facial Diagnosis,” in IEEE Access, vol. 8, pp. 123649-123661, 2020. “Improvement of generalization ability of deep CNN via implicit regularization in two-stage training process,” IEEE Access, vol. 6, pp. 15844-15869, 2018.

Response:

We thank the reviewer for the comment. The above-stated comment is discussed and referenced in the first section as follows:

Reinforcement learning evaluates a policy for the given action-value function based on the behaviour observed. In the training process, regularization has an important role in adjusting the hyperparameters such as the environment for analyzing different scenarios [44]. For improving the generalizability of the learning method, efficient regularization needs to be designed. Regularization is required in complex environments to overcome the problems such as overfitting during the training phase [45-47]. In the proposed work, the environment is simple comprising dog and owner, hence there will be lower complexity.

Round 2

Reviewer 1 Report

It can be accepted.

Reviewer 2 Report

All comments have been carefully revised, so this paper is recommended for acceptance.