Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
This paper proposes a less principled but more successful alternative routing mechanism for capsule networks. It replaces the EM or dynamic algorithm with a simple perceptron model for predicting routing coefficients. This routing mechanism is thoroughly explored for its test performance, transformation robustness, as well as parameter and computational requirements. The results are convincing. The mechanism they propose improves test performance and robustness to adversarial attacks and view point transformations without significantly increasing parameters or computational cost. the paper is original. The routing procedure they propose is simple and somewhat self evident, but the improved performance they find as a result of the mechanism is impressive. And i know of no other work that has attempted to replace the routing mechanism with such a technique. The paper is well written though some elements could be made more clear (detailed in the improvements section) The paper is fairly significant as it explores a baseline for routing that i believe many researchers may have written off for being to simple, but demonstrates its success convincingly.
Post-rebuttal: I have considered the opinion and viewpoint of the other reviewers, who have both provided some good insight on the paper. I have also read the response of the authors very carefully, which has provided some more information. I am happy to revise my score reflecting the new evidence authors have provided. --------------------------------- ---Authors propose a new routing algorithm for CapsNets called Self-routing which is a supervised non iterative routing method that removes the agreement component from the routing process. Authors base their implementation on mixture of experts claiming a resemblance between a capsule and an expert. In that sense an expert is specialising in a different region of the input space, whose contributions are adjusted differently per example/input. What happens in the Dynamic Routing and EM is that the agreement between a higher level and lower level capsule is paramount for deciding if something is present in an image or which information to keep based on a voting process. This component is happening in an unsupervised manner. Authors claim that this property might be attributed to an ensemble (averaging) behaviour and not to the routing-by-agreement per se. I think this might require more experimental results than those presented in this paper. ---This paper presents some original results in terms of giving a different spin on the routing algorithms for CapsNets but it does not provide compelling arguments as to why this might be happening. Primarily I think the information pertaining to the how self-training process actual works in practise is rather limited, other than the fact that a pose vector is multiplied by a trainable weight matrix and outputs the routing coefficients. Maybe section 4.2 could include a step-by-step description of the process rather than relying on the mixture-of-experts process alone. ---The paper is clearly written but the different tables presenting various results take time to understand what they are referring to, e.g. parameters, error rates, etc. The experiments are comprehensive focusing on three different datasets and also the use of adversarial attacks. However some clarity on the point made in line 60 would be helpful. As far as I am aware neither dynamic routing nor EM routing papers mentions anything about using ResNet as part of their experiments with CapsNets other than defining a CNN baseline. Could authors elaborate a bit more on that? Is it the case they implemented Dynamic Routing and EM from scratch and then added ResNet in all implementations to enable comparisons with CIFAR-10 etc.? Is there a reason why the employed a CNN to do so instead of relying solely on CapsNets? Do they suggest that as it stands at the moment the best way forward is an ensemble of ResNet type models and CapsNets. ---The arguments and the quality of the work presented seem plausible to a large extent - but more elaboration on section 4.2 would help to improve the theory behind the concept of the new routing algorithm. ---CapsNets are a very promising area of research therefore new results, methods and insights are significant. The extend to which the methods presented in this paper are significant or novel is limited, but have an important spin towards changing the routing process not to include the agreement component. Could you please check the paper once more for correcting typos, e.g. robust instead of robustly in line 16, line 42 much more, etc.
Post rebuttal : The authors addressed the concerns I raised regarding error bars in the experiments. I am happy to update my rating to 7. -------------- This paper proposes a self-routing mechanism in which each capsule's pose goes through a little neural net which outputs the routing coefficients to each potential parent capsule. This is in contrast to previous work, where a key component of the model design was that routing of parts to wholes should be determined based on agreement of the pose of the whole across parts. The obvious drawback seems to be that if a part can reasonably belong to multiple wholes, it does not get the chance to choose one based on what other parts are agreeing on. Instead it must make a choice a-priori or just spread its vote across the many choices. However, it turns out that this is not detrimental, and in fact is advantageous, at least on tasks such as classification, adversarial robustness, and generalization to novel view points. Pros - The paper questions an important idea behind capsule networks. This helps answer the question whether it is routing by agreement that is important or just routing. - Experiments cover 3 different tasks. Cons - The results don't have error bars and based on all the numbers it seems that the improvement in the results may not be very significant. However, even matching the results is enough to make the point. Overall, the paper is interesting since it provides evidence that strong gating alone is an important factor that the community should pay attention to when designing variants of capsule networks. Alternatively, this suggests that the tasks that are being used for evaluating capsule networks are not making use of the agreement aspect of the routing. Either way the conclusions are interesting.