News from the MLSP TC

By Kenneth E. Hild II (MLSP Liaison to E-Newsletter)

MLSP workshop

The 2012 MLSP Workshop, which was held in The Royal Palace of the Magdalena in Santander, Spain, concluded on September 26. This was the 22nd annual meeting organized by the MLSP TC and was a very successful event. The meeting featured three plenary talks. The plenary speakers, which include two speakers from the Machine Learning and NIPS community, were: Prof. Martin J. Wainwright of UC Berkeley (Learning and Message-Passing in Graphical Models), Dr. Francis Bach from INRIA, France (Large-scale convex optimization for machine learning), and Prof. Ali H. Sayed from UCLA (Adaptation and Learning over Complex Networks). We received a total of 202 submissions, of which 108 were accepted for presentation (53% acceptance ratio). The distribution of accepted papers was as follows: Learning Theory 11%, Bayesian 13%, Speech/Audio 11%, Blind Source Separation 7%, Classification 23%, Image/Video 14%, and Other 21%. The technical program can be found on the workshop website. A total of 124 people registered for the workshop. The registrants were mainly from Europe (60%), USA (29%), and Asia (7%). We would like to thank the following for supporting this year's workshop: Ayuntamiento de Santander, Gobierno de Cantabria Consejeria de Educacion Cultura y Deporte, Amazon Information Security, Pascal 2, Universidad de Cantabria, Universidad Carlos III de Madrid, Asociacion Espanola Ingenieros de Telecomunicacion Cantabria, Gobierno de Espana Ministerio de Ciencia E Innovacion.

The 2013 MLSP Workshop will be held in Istanbul, Turkey, and the 2014 MLSP Workshop will be held in Reims, France. Typically, the paper deadline is in March or April and the meeting is held in September. Be sure to mark your calendars for these two upcoming events.

Data Analysis Competition

For each of the last eight MLSP Workshops, the MLSP TC has hosted a Data Analysis Competition. The most recent competition required the entrants to select/design a classifier and any necessary pre-processing system to correctly classify an employee/resource pair (a resource can be, e.g., a computer, data base, or data portal) into one of two classes: access or non-access. The winner was determined as the submission that minimized the number of manual access grant/removal operations in the disjoint and unpublished testing data. The winning team, the members of which are affiliated with the Indian Institute of Technology, Kanpur, India, consists of: Ankit Gupta, Shashwat Mishra, and Amitabha Mukerjee. The second-place team, the members of which are affiliated with Tampere University of Technology, Tampere, Finland, consists of: Heikki Huttunen, Timo Erkkila, Pekka Ruusuvuori, and Tapio Manninen. The winning teams received Kindle Fire tablet computers from Amazon, free registration to the MLSP Workshop, and a free trip to present their results at Amazon's new headquarters in Seattle, Washington. We wish to congratulate the winners on their achievement.

New members

The MLSP TC recently held an election for membership in the MLSP TC. The term of four of the current TC members expires in December. Three of these members are seeking re-election. In addition, ten new members were nominated. Thirteen total nominees are competing for seven open slots. The results of the voting will be available around mid-November and new members will serve from January 1, 2013, to December 31, 2015. Those who are interested in joining the MLSP TC in the future should read the information on the MLSP TC website.

Quality of Peer Reviewing

At the most recent MLSP TC business meeting in Santander, the TC discussed a number of topics. In addition to workshop reports (past, present, and future), we discussed how to increase the participation from the machine learning community, the free access option for the MLSP proceedings, options for new transactions, and other outreach activities. We also had a discussion on how to improve the quality of peer reviews. On the latter topic, Tulay Adali led an effort to discern how to improve the quality of reviews for ICASSP and ICIP and for other SPS-sponsored workshops. She worked directly with researchers from other TC's including, most notably, Gaurav Sharma, John Hansen, and Ken Sugiyama. The main points that resulted from the discussion with the review panel, which resulted in a set of recommendations to the Conference Board, include:

• build memory into the database for future reference by the TC/area chairs (reviewers will be graded based on the usefulness and the timeliness of their reviews; reviewers might be classified as senior and regular so that at least one senior reviewer can be included on each paper, as was done for ICIP 2012 and AASP);
• increase the list of potential reviewers by, e.g., reaching out to Affiliate members (the MLSP TC has, e.g., asked all Affiliate members to send a list of their most relevant publications, if they are interested in reviewing for ICASSP 2012);
• provide feedback to reviewers (reviewers are likely to participate better if they have stronger sense of being a part of the community and do not feel that they are simply being employed as laborers; we could communicate to reviewers the decisions that were made on the papers they reviewed, as well as include responses from the authors);
• require a second round of reviews (in which case we could, e.g., check if the reasonable comments of the reviewers were incorporated; however, the implementation might be difficult due to timing of the decisions and because of the associated increase in workload for the TC/area chairs) or, at the very least, provide a place for the authors to respond to reviewer comments when they upload the final version of the paper;
• improve the review form (e.g., the instructions to the reviewers should emphasize that justification for their evaluation is required irrespective of the recommendation, which is especially helpful for marginal papers; review forms for ICASSP 2012 required inclusion of comments only if the reviewer gave a low score and was not required for marginal scores, which resulted in an increase in the number of marginal papers and a decrease in the number of helpful comments; improving the review form decreases the load on the TC/area chairs since a clear explanation of the reasons for the rejection must be provided for rejected papers);
• select and promote a single weighting scheme to be used for paper acceptance (currently two different weighting schemes are used, one based on acceptance and one based on the overall scores in different categories; the two methods result in very different rankings and selecting just one might provide some level of balance in how the decisions are performed across TCs);
• recognize good reviewers (e.g., at ICIP 2012, the plan is to recognize a small number of outstanding reviewers by giving them a recognition certificate and a banquet ticket), and
• ensure the list of topics (usually adopted from the TSP EDICS) adequately represent the TC.

For anyone who is interested in following MLSP activities, we encourage you to sign up for the MLSP email list. The invitation can be found at the top of the MLSP TC website.