Wednesday, 9 July 2014

WCCI 2014 Day 3

The first event of today was a plenary panel session organised by Nikhil Pal: Is "Publish or Perish" causing "Death" of science?

Nikhil described the academic system as being like a pyramid, with a very few at the very top and many at the bottom. With so many (too many, in his opinion) at the bottom, there is a tremendous pressure to publish more, thanks in part to an emphasis on metrics like impact factor and h-index. This has resulted in a proliferation of conferences and journals, and an increase in publishing misconduct like plagiarism.

Hisao Ishibuchi gave a thoughtful analysis on the implications of quantity vs quality. His analysis looked at the impact each dimension has on an academic's evaluation by administrators.

Chin-Tong Lin spoke about the victims of the pressure caused by Publish or Perish. The main victims are Associate Editors, and Editors in Chief, who all have their workloads increased by an increase in submissions. He demonstrated how, for the journal IEEE Transactions on Fuzzy Systems, the increase over time of impact factor for the journals led to an increase in submissions. Other victims were the readers and authors of journals: the readers who are subjected to more low-quality papers and authors who must produce the articles.

Derong Liu told the audience that the pressure affected quality "moderately" and increased plagiarism "sort of". He didn't think that the "pay to publish" approach would improve publication quality, and that metrics do influence publications.

Simon Lucas offered his opinion that while the average quality of papers may be going down, the average quality of papers being read is going up. Authors can get sucked into publishing too  much quantity, which negatively affects quality and encourages plagiarism. The "pay to publish" approach will give an advantage of wealthy authors but can be made to work. He said that impact factor must be taken seriously, as it does inflluence publications. In his opinion, the publishing system isn't broken but does need to include more social media in the publications process, especially discussions on articles.

Marios Polycarpou spoke next and told the audience that researchers now spend more time reporting research than actually doing it. This has caused the signal to noise ratio to go down due to the pressure to publish. This pressure comes from too much emphasis on formulas such as h-index, impact factor and so on. Technology makes it so easy to detect plagiarism that it just isn't worth the risk, since one case of plagiarism can ruin a researcher's career.

Jun Wang presented data that showed that the emphasis on pubishing has promoted fraudulent research, and that impact factors are causing coercive citations, that is, citations that authors are coerced into including to boost the journal impact factor. Impact factors also have a strong correlation with article retractions.

Xin Yao was the last speaker and was to the point: trying to quantify something that is not quantifiable is a symptom of lazy management. The pressure to publish has a tremendous impact of younger researchers, which leads them to take risks with plagiarism and academic fraud.

Yann LeCun, Director of AI at Facebook, gave a plenary on "Deep Learning and the Representation of Natural Data". The third renewal of neural nets has focused mostly on audio and video data. An overview of deep learning explained the case for deep learning over SVM approaches, which are effectively lookup/template methods. Various example applications were discussed, such as face detection and body pose recognition. The most exciting part was the array of demos! A video was shown of a convolutional network taking images from a Kinect and constructing a 3D model of a hand to idenitfy gestures. Yann plugged in a standard webcam and feed the images into a trained convolutional network, and the webcam recognised a variety of objects in the hall, such as Yann's laptop, keyboard, iPod (well, iPhone, but close enough!) sunglasses, coffee mugs etc. Another webcam demo with a convolutional network demonstrated how it can learn a variety of objects by pointing the webcam at an object and clicking 'learn'. The final video showed Yann's students using a pair of stereo vision cameras with a convolutional network to identify traversible and non-traverisble terrain for a robot to navigate.
 

No comments:

Post a Comment