In my last blog post, I discussed about the contents of the paper called Going Deeper with Convolutions written by the authors (Szegedy et al., 2014). In this post, I am going to give a review of the paper according to the format of IJCAI (International Joint Conferences on Artificial Intelligence Organization).
Relevance (1-10 stars)
10 Starts: This paper fully meets the scope of conference.
Significance (1-10 stars)
9 starts: The paper represents new state of the art in large-scale image classification by increasing the networks width and height while keeping the computational expense constant. This paper would be an ideal knowledge source for people who want to gain more insights about deep architecture.
Originality (1-10 stars)
9 stars: The building blocks of the image classification model that the authors present is based on the reduced dimensions Inception module which is built on top of Naïve Inception module. Therefore, I believe this paper stands upon original findings.
Technical Quality (1- 10 stars)
9 stars: This paper achieves the highest accuracy in terms of the performance among other competitors in the ImageNet classification challenge. Therefore, the result is technically best.
Clarity and quality of writing (1-10 stars)
7 stars: The images and the tables used in this paper are clearly illustrated and explained. The choice of vocabulary seems really academic and professional. However, there exist various tables with numeric values, which seems confusing.
Scholarship, i.e. scientific context (1-10 stars)
8 stars: This paper is not just about improving accuracy to a specific task but to introduce a new model which performs better in terms of time and computational expense. Therefore, it can clearly be stated that this paper is a good example of a scientific and scholarly work.
Overall Score: The overall score of this paper is 8.67 based on my assessment in IJCAI criteria. Therefore, this paper falls under the category of a very strong paper and the status is accepted. This paper stood first in the ImageNet classification challenge.
Confidence of my assessment
I would rate myself 6.5 out of 10. Looking and analyzing the paper several times I have gained modest knowledge in this domain to assess the work. However, there might be other advanced literatures published in this area where my level of understanding would be lesser.