Relevance: 6/10
This paper is more close to control engineering than to main AI topics. The neural network is used, but not in way like deep learning or reinforcement learning. It is not clear if this work is interesting for many AI researchers. If author can compare with more AI method, it will help.
Significance: 8/10
The idea of self-organizing RBF network with spiking neuron mechanism is interesting. I think this is good for control system which is nonlinear and changing. The example of wastewater plant is practical and realistic. It show the method is working.
But it is only one example. No test in other systems. Also, the paper not say if the method can be used for more fast or real-time system.
Originality: 7/10
Author combine many known idea like RBF, recurrent NN, Lyapunov and spiking to make one model. The growing and pruning of hidden neuron is done with new idea. This part I think is creative.
But I not understand why this growing method is better than others. Why use this formula? Why the threshold like 0.05? Author should explain this more.
Technical Quality: 7/10
The math in the paper is correct, and I think the stability proof is good. The simulation looks fine. The paper has strong theory.
But no test for different type of systems. No baseline from new AI methods like LSTM, GRU, or RL. Also, author do not say how fast the model run or how much memory it use. This is important for control in real world.
No ablation study too. It is not shown if the self-organizing part really help or not.
Clarity and Quality of Writing: 6/10
The paper is not easy to read. Sentences are long and many equations. Too much math and not enough explanation. Some figures are small or confusing. For example, Fig. 3 and Fig. 4 not explain enough.
Author should write more simple. Also good to add more picture, or one example with small system to understand better.
Scholarship (Scientific Context): 7/10
Many citations in paper, but most are old (before 2015). No mention of deep learning for control, like using LSTM or reinforcement learning. These are very active area in AI now. Author should include some of this in related work and discussion.
Overall Score: 7/10 — Accept
I think the paper is strong for theory and has good idea. But it needs better writing and more comparison with modern AI. I think it can be better if authors improve this.
For now, I recommend accept, but not strong accept.
Confidence in Assessment: 6/10
I know model predictive control and Waste Watertreatment Plants also, but it is not my main topic.
Comments to Authors:
-
Please show how your method is different or better than deep learning method, like LSTM-MPC or RL-based controller.
-
Explain better why you use this special spiking neuron function. Why log and sine in the equation? Why this threshold?
-
Add figure to show how network change over time (more or less neurons).
-
Say something about run time of training and inference. Is this real-time?
-
Writing can be more easy. Try to write with short sentences and more picture.
-
Update your reference list with papers from 2016 to now.
The blog post is clear for me to read with the actual scores highlighted. It also provides for each category a clear reasoning and suggestions. I really liked the honest confidence assessment and the comments to the authors were also really helpful/constructive, especially asking them to compare it with modern AI methods - overall great job!
AntwortenLöschen