Ever wondered if we could build a deep learning neural network which can predict whether your research paper will be submitted or not in the conference? If yes, then I am sure this blog will be an interesting read.
I was recommended an interesting paper a couple of months ago and the following blog will be more of a summary of the paper – https://arxiv.org/pdf/1812.08775.pdf
With the increase in the deep learning domain, we have seen a surge in the number of papers which are submitted in the conferences. Traditionally, “peer review” system was in place to select/reject the paper. However, the authors of this paper have discussed a deep convolutional network approach which can help in automating the process.
One has to keep in mind that the author of this paper – Jia-Bin Huang has taken the “gestalt” way. It basically means that he is classifying the papers just by the way they look! After reading this I am sure that you are jumping on your seats to try out a nlp based approach, I’d suggest to be patient and keep reading 🙂 .
What did they achieve?
Quoting directly –
Trained on ICCV/CVPR conference and workshop papers from 2013 – 2017, our deep network based classifier achieves 92% accuracy on papers in CVPR 2018. Our model safely rejects the number of bad paper submissions by 50% while sacrificing only 0.4% of good paper submissions. Our system can thus be used as a pre-filter in a cascade of the paper review process.
The paper says that they crawled this website and gathered the conference and workshop papers from 2013-2017. They only downloaded the papers with more than 7 pages.
The pdf2image python wrapper was used to convert the papers to images. The original size of the converted image is of size 2200 × 3400 pixels. The images were then resized to 224 × 224 pixels for both training and testing.
The data preprocessing is interesting, you should read about that in paper!
Deep Neural Network Architecture –
ResNet architecture was used.
Just the last layer which has 1000 nodes in the picture(because of imageNet dataset) was replaced by 2 output nodes. (transfer learning!)
This ImageNet pre-trained network was fine tuned with stochastic gradient descent (SGD) with a momentum of 0.9 for a total of 50 epochs. The learning rate was set to 0.001 and it was decayed by a factor of 0.1 after every 10 epochs. The loss function was weighted cross-entropy loss.
Using GANS to generate good CVPR papers?
The author has used progressive growing of GANs technique to generate papers which is fun to look – https://www.youtube.com/watch?v=yQLsZLf02yg&feature=youtu.be
The paper itself lays out critically the drawbacks and limitations which I request you to read. I am mentioning the obvious ones –
- good materials but bad visual layout may get rejected or accept crappy papers with good layout may get accepted!
- both the classifier and the generative model assume that all the papers have the same typesetting style.
Things to ponder –
After reading the paper(or this blog) if you are wondering what to do next then here are a few tips for you
- Try building nlp based classifier which checks for content and not just “paper gestalt”.
- Try using Diverse image-to-image translation via disentangled representations approach to generate diverse good paper visuals
- Lastly, read the paper! 😉
I hope this was helpful, feel free to ping me if you have any doubts or if you want to catch over coffee!
Happy Coding! 🙂