The idea of a symposium to bring together the bright new talent within the SIGMM community and to hear their views on some topics within the area and on the future of Multimedia, was first mooted in 2014 by Shih-Fu Chang, then SIGMM Chair. That lead to the “Rising Stars Symposium” at the MULTIMEDIA Conference in 2015 where 12 invited speakers made presentations on their work as a satellite event to the main conference. After each presentation a respondent, typically an experienced member of the SIGMM community, gave a response or personal interpretation of the presentation. The format worked well and was very thought-provoking, though some people felt that a shorter event which could be more integrated into the conference, might work better.
For the next year, 2016, the event was run a second time with 6 invited speakers and was indeed more integrated into the main conference. The event skipped a year in 2017, but was brought back for the MULTIMEDIA Conference in 2018 and this time, rather than invite speakers we decided to have an open call with nominations, to make selection for the symposium a competitive process. We also decided to rename the event from Rising Stars Symposium, and call it the “SIGMM Emerging Leaders Symposium”, to avoid confusion with the “SIGMM Rising Star Award”, which is completely different and is awarded annually.
In July 2018 we issued a call for applications to the “Third SIGMM Emerging Leaders Symposium, 2018” which was to be held at the annual MULTIMEDIA Conference in Seoul, Korea, in October 2018. Applications were received and were evaluated by a panel consisting of the following people, and we thank them for volunteering and for their support in doing this.
Werner Bailer, Joanneum Research
Guillaume Gravier, IRISA
Frank Hopfgartner, Sheffield University
Hayley Hung, Delft University, (a previous awardee)
Marta Mrak, BBC
Based on the assessment panel recommendations, 4 speakers were included in the Symposium, namely:
Hanwang Zhang, Nanyang Technological University, Singapore
Michael Riegler, Simula, Norway
Jia Jia, Tsinghua University, China
Liqiang Nie, Shandong University, China
The Symposium took place on the last day of the main conference and was chaired by Gerald Friedland, SIGMM Conference Director.
Towards X Visual Reasoning
By Hanwang Zhang (Nanyang Technological University, Singapore)
For decades, we are interested in detecting objects and classifying them into a fixed vocabulary of lexicon. With the maturity of these “low-level” vision solutions, we are hunger for a “higher-level” representation of the visual data, so as to extract visual knowledge rather than merely bags of visual entities, allowing machines to reason about human-level decision-making. In particular, we wish an “X” reasoning, where X means eXplainable and eXplicit. In this talk, I first reviewed a brief history of symbolism and connectionism, which alternatively promote the development of AI in the past decades. In particular, though the deep neural networks — the prevailing incarnation of connectionism — have shown impressive super-human performance in various tasks, they still lag behind us in high-level reasoning. Therefore, I propose the marriage between symbolism and connectionism to take the complementary advantages of them, that is, the proposed X visual reasoning. Second, I introduced the two building blocks of X visual reasoning: visual knowledge acquisition by scene graph detection and X neural modules applied on the knowledge for reasoning. For scene graph detection, I introduced our recent progress on reinforcement learning of the scene dynamics, which can help to generate coherent scene graphs that respect visual context. For X neural modules, I discussed our most recent work on module design, algorithms, and applications in various visual reasoning tasks such as visual Q&A, natural language grounding, and image captioning. At last, I visioned some future directions towards X visual reasoning, such as using meta-learning and deep reinforcement learning for more dynamic and efficient X neural module compositions.
Professor Ramesh Jain mentioned that a truly X reasoning should consider the potential human-computer interaction that may change or digress a current reasoning path. This is crucial because human intelligence can reasonably respond to interruptions and incoming evidences.
We can position X visual reasoning in the recent trend of neural-symbolic unification, which gradually becomes our consensus towards a general AI. The “neural”’ is good at representation learning and model training, and the “symbolic” is good at knowledge reasoning and model explanation. One should bear in mind that the future multimedia system should take the complementary advantages of the “neural-symbolic”.
BioMedia – The Important Role of Multimedia Research for Healthcare
by Michael Riegler (SimulaMet & University of Oslo, Norway)
With the recent rise of machine learning, analysis of medical data has become a hot topic. Nevertheless, the analysis is still often restricted to a special type of images coming from radiology or CT scans. However, there are continuously vast amounts of multimedia data collected both within the healthcare systems and by the users using devices such as cameras, sensors and mobile phones.
In this talk I focused on the potential of multimedia data and applications to improve healthcare systems. First, a focus on the various data was given. A person’s health is contained in many data sources such as images, videos, text and sensors. Medical data can also be divided into data with hard and soft ground truth. Hard ground truth means that there are procedures that verify certain labels of the given data (for example a biopsy report for a cancerous tissue sample). Soft ground truth is data that was labeled by medical experts without a verification of the outcome. Different data types also come with different levels of security. For example activity data from sensors have a low chance to help to identify the patient whereas speech, social media, GPS come with a higher chance of identification. Finally, it is important to take context into account and results should be explainable and reproducible. This was followed by a discussion about the importance of multimodal data fusion and context aware analysis supported by three example use cases: Mental health, artificial reproduction and colonoscopy.
I also discussed the importance of involving medical experts and patients as users. Medical experts and patients are two different user groups, with different needs and requirements. One common requirement for both groups is the need for explanation about how the decisions were taken. In addition, medical experts are mainly interested in support during their daily tasks, but are not very interested in, for example, huge amounts of sensor data from patients because the increase amount of work. They have a preference on interacting with the patients than with the data. Patients on the other hand usually prefer to collect a lot of data and get informed about their current status, but are more concerned about their privacy. They also usually want that medical experts take as much data into account as possible when making their assessments.
Professor Susanne Boll mentioned that it is important to find out what is needed to make automatic analysis accepted by hospitals and who is taking the responsibility for decisions made by automatic systems. Understandability and reproducibility of methods were mentioned as an important first step.
The most relevant messages of the talk are that the multimedia community has the diverse skills needed to address several challenges related to medicine. Furthermore, it is important to focus on explainable and reproducible methods.
Mental Health Computing via Harvesting Social Media Data
By Jia Jia, Tsinghua University, China
Nowadays, with the rapid pace of life, mental health is receiving widespread attention. Common symptoms like stress, or clinical disorders like depression, are quite harmful, and thus it is of vital significance to detect mental health problems before they lead to severe consequences. Professional mental criteria like the International Classification of Diseases (ICD-10 ) and the Diagnostic and Statistical Manual of Mental Disorders (DSM ) have defined distinguishing behaviors in daily lives that help diagnosing disorders. However, traditional interventions based on face-to-face interviews or self-report questionnaires are expensive and hysteretic. The potential antipathy towards consulting psychiatrists exacerbates these problems.
Social media platforms, like Twitter and Weibo, have become increasingly prevalent for users to express themselves and interact with friends. The user-generated content (UGC) shared in such platforms may help to better understand the real-life state and emotion of users in a timely manner, making the analysis of the users’ mental wellness feasible. Underlying these discoveries, research efforts have also been devoted for early detection of mental problems.
In this talk, I focused on the timely detection of mental wellness, focusing on typical mental problems: stress and depression. Starting with binary user-level detection, I expanded the research by considering the trigger and the severity of the mental problems, involving different social media platforms that are popular in different cultures. I presented my recent progress from three prespectives:
- Through self-reported sentence pattern matching, I constructed a series of large-scale well-labeled datasets in the field of online mental health analysis;
- Based on previous psychological research, I extracted multiple groups of discriminating features for detection and presented several multi-modal models targeting at different contexts. I conducted extensive experiments with my models, demonstrating significantly better performance as compared to the state-of-the-art methods; and
- I investigated in detail the contribution per feature, of online behaviors and even cultural differences in different contexts. I managed to reveal behaviors not covered in traditional psychological criteria, and provided new perspectives and insights for current and future research.
My developed mental health care applications were also demonstrated in the end.
Dr. B. Prabhakaran indicated that mental health understanding is a difficult problem, even for trained doctors, and we will need to work with psychiatrist sooner than later. Thanks to his valuable comments, regarding possible future directions, I envisage the use of augmented / mixed reality to create different immersive “controlled” scenarios where human behavior can be studied. I consider for example to create stressful situations (such as exams, missing a flight, etc.), for better understanding depression. Especially for depression, I plan to incorporate EEG sensor data in my studies.
Towards Micro-Video Understanding
By Liqiang Nie, Shandong University, China
We are living in the era of ever-dwindling attention span. To feed our hunger for quick content, bite-sized videos embracing the philosophy of “shorter-is-better”, are becoming popular with the rise of micro-video sharing services. Typical services include Vine, Snapchat, Viddy, and Kwai. Micro-videos like a wildfire are very popular and taking over the content and social media marketing space, in virtue of their value in brevity, authenticity, communicability, and low-cost. Micro-videos can benefit lots of commercial applications, such as brand building. Despite their value, the analysis and modeling of micro-videos is non-trivial due to the following reasons:
- micro-videos are short in length and of low quality;
- they can be described by multiple heterogeneous channels, spanning from social, visual, and acoustic to textual modalities;
- they are organized into a hierarchical ontology in terms of semantic venues; and
- there are no available benchmark dataset on micro-videos.
In my talk, I introduced some shallow and deep learning models for micro-video understanding that are worth studying and have proven effective:
- Popularity Prediction. Among the large volume of micro-videos, only a small portion of them will be widely viewed by users, while most will only gain little attention. Obviously, if we can identify in advance the hot and popular micro-videos, it will benefit many applications, like the online marketing and network reservation;
- Venue Category Estimation. In a random sample over 2 million Vine videos, I found that only 1.22% of the videos are associated with venue information. Including location information about the videos can benefit multifaceted aspects, such as footprints recording, personalized applications, and other location-based services, it is thus highly desired to infer the missing geographic cues;
- Low quality sound. As the quality of the acoustic signal is usually relatively low, simply integrating acoustic features with visual and textual features often leads to suboptimal results, or even adversely degrades the overall quality.
In the future, I may try some other meaningful tasks such as micro-video captioning or tagging and detection of unsuitable content. As many micro-videos are annotated with erroneous words, namely the topic tags or descriptions are not well correlated to the content, this negatively influences other applications, such as textual query search. It is common that users upload many violence and erotic videos. At present, the detection and alert tasks mainly rely on labor-intensive inspection. I plan to create systems that automatically detect erotic and violence content.
During the presentation, the audience asked about the datasets used in my work. In my previous work, all the videos come from Vine, but this service has been closed. The audience wondered how I will build the dataset in the future. As there are many other micro-video sites, such as Kwai and Instagram, I hence can obtain sufficient data from them to support my further research.