2. Executive summary
New South Wales Public Health Bulletin · 24.06.2010 · 21 (1) · 9-10
DOI: 10.1071/nb10s02 ↗Kurz gefasst
Fachartikel in der Zeitschrift New South Wales Public Health Bulletin, online erschienen am 24.06.2010.
Abstract (Originalsprache)
ion Challenges: The major challenge in this area continues to be the tension between day-to-day work and abstractions -in developing the foundations and the theory, one must work at higher-level abstractions, but for solving a specific problem in a specific domain, those abstractions may be less useful and specialized solutions might be more desirable. Another perspective is that we need to find the right abstractions that can capture the observed behavior of algorithms in specific domains. Data Access Challenges: Navigating data access can be quite tricky -without the right kind of data, it is difficult to understand what are the most effective and crucial foundational questions to focus on. A lot of researchers focus on what is called "Internet" problems, i.e., problems faced by big Internet companies like Google and Facebook. While those problems are important, more work needs to be done on hard questions in physical sciences, where it is also easier to get access to data. However, the challenge here is that interpretation of the data and results requires specific domain expertise and access to collaborations. Communication Obstacle: Finally, communication between the researchers working on foundations and researchers working in application domains continues to be an issue, and raises significant barriers in developing new theories that may be applicable to emerging application domains. 5.5.3 Strategic Priorities Some of the major open questions in this area also constitute the main strategic research priorities. 1. Although there is much work on understanding tradeoffs between different resources, research needs to be supported to improve our understanding and propel the development of better techniques that can efficiently navigate a large number of competing resource considerations. For example, when doing inference on large amounts of high-dimensional data, how can we balance the desire for a few samples, low error, and efficient computation? 2. Research projects need to be supported that focus on developing algorithms, especially machine learning algorithms, for dealing with a large number of dimensions (features) and a small number of samples (e.g., in drug testing, or cancer research). How can one design algorithms and statistical methods in such a scenario and reason about the accuracy of those? A related challenge is that of doing inference in presence of a large number of labels, e.g., in personalization. Validation is also nontrivial when there is limited access to fresh data. We need methodologies to think about this question properly, and for doing evaluation when constrained by the amount of annotated data available. 3. Initiatives focused on understanding the consequences of complex human-machine interaction systems need to be supported across agencies. In systems where there are a large number of interacting agents, some of which may be computers and some human, there may be unintended consequences (e.g., discriminatory automated policies). 4. An important direction of research is the development of practicable theory of "causality". Although there has been much work on this topic, the practical impact of that work has been low. Funding both theoretical and development work in this area, as a partnership is an important direction. 5.6 Large-scale Inference & Learning 5.6.1 Current Landscape There has been a tremendous interest in extracting insight from large-scale data sets collected from, for example, users’ input, sensor readings, and scientific experiments. To perform high-performance inference and learning from these data sets is a nontrivial task, especially with constraints on computing and storage resources exist. To address this, the community has been exploring and developing novel methodological approaches to parallel and distributed computation for learning and inference. At the core, there has been significant progress in the fields of databases, high-performance computing, and hardware towards supporting large-scale stochastic optimizations. Notable examples include large-scale logistic regression and deep learning, which has benefited greatly from the development of new architectures, algorithms, and protocols that leverage the computing power of specialized hardware such as GPU and FPGAs. These advances in the science of large-scale inferences and learning have led to successes in a wide range of application domains: computer vision, speech recognition, and natural language processing have been greatly improved by adopting deep learning; cancer genomics has made modest progress at a scale that it hasn’t been able to do before. 5.6.2 Major Challenges and Obstacles This section discusses some of the many challenges and obstacles associated with large-scale inference and learning. We group these challenges into three areas: data availability, bias and noise, and domain specific vs. domain agnostic challenges. Data Availability Challenge: Academic data sets vary in their sizes and quality, but, as of May 2016, there is a lack of large enough data sets or benchmarks that are widely adopted/accepted by the community to evaluate the effectiveness and performance of proposed algorithms or systems. Those data sets would be easier to get from industry; however, industry either does not have the incentive or cannot release their datasets because of privacy concerns. It would be tremendously helpful to facilitate a datasharing platforms and better partnerships between industry and academia. Bias and Noise Challenges. Even if one has access to a large-scale data set, often the data are still largely unlabeled, which creates challenges for designed inference or learning algorithms to understand the bias and noise in the data set. For example, the data collection process may reflect only a skewed perspective of the data, loss of accuracy may be introduced due to flawed data collection, etc. Domain-specific vs. Domain-agnostic Challenges. Traditional learning algorithms such as SVMs are not necessarily optimized for certain domain-specific problems (e.g., in large-scale scientific computation), which leads to low performance or accuracy. Computer scientists have come up with better domain-specific solutions; we have seen many high-profile successful applications in the fields of speech recognition, computer visions, and computational genomics. However, there is still a gap between the advances in computer science and the awareness of these advances in a domainspecific context. An urgent effort is needed to help people of all disciplines, who are non-experts in computer science, know what is available to them. 5.6.3 Strategic Priorities The breakout sessions considered the following areas as strategic priorities for advancing the innovation in large-scale inference and learning, as well as broadening understanding of the field. 1. Though there have been many high-profile successes in developing platforms dedicated for specific inference and learning tasks, it is still important to give priorities to combining “systems” (such as databases, HPC, hardware) and ML algorithms: fundamental progress in systems research may bring the inference and learning algorithms to a new level. For example, neural networks were proposed over two decades ago, but didn’t get much traction not only because of the unavailability of large-scale data, but also the lack of computing platforms that can support its demand of computing resources. 2. Machine learning has not evolved into a state where machine can automatically decide “the best” algorithm for specific inference and learning tasks on a specific dataset; currently, researchers need to experiment and choose from different algorithms to find the best algorithm and parameters. It is important to invest in research that advances the human-machine interactions for inference/learning, such as in mitigating the impact of bias and noise in data set. 3. Data driven inference and learning is starting to make decisions that humans cannot understand; for example, AlphaGo makes excellent moves that even human champions find hard to understand. Thus, another strategic priority is to make machine learning interpretable, e.g., in terms of knowledge of the processes generating the data, such that human not only receives the final result of inference and learning, but also the intelligence and strategies that lead the learning to the result. 4. The training of non-experts in computer science is not prevalent in the graduate programs of different disciplines. For example, computational skills are not taught in a graduate class in biology; people have to self-teach themselves. Researchers outside of CS have difficulties in accessing and understanding these algorithms because they are written in a way that assumes significant background in computer science. It is import to develop classes to get non-CS researchers and students to understand what CS people are capable of making for them. 5.7 Learning Analytics and Education 5.7.1 Current Landscape There has been an intense amount of work in recent years on improving learning and education using data-driven methods, and this remains one of the most important application areas for big data. Learning analytics is instrumental in supporting innovations in education, including individual and collaborative learning environments, MOOCs, intelligent tutors, learning management systems, open-ended learning environments, adaptive hypermedia, flipped classrooms, learning at scale, etc. It is also a key to personalized education, and through predictive modeling can help tailor education to the needs of a diverse student body. Much of the work in this area to date has focused on developing mechanisms to generate and collect assessment reports, e.g., through use of web-based intelligent tutors. Massively Open Online Courses (MOOCs) naturally generate large volumes of data, and there has been much work on analyzing that data, especially for stu
Zitieren
APA
(2010). 2. Executive summary. New South Wales Public Health Bulletin, 21(1), 9-10. https://doi.org/10.1071/nb10s02
Vancouver
2. Executive summary. New South Wales Public Health Bulletin. 2010;21(1):9-10. doi:10.1071/nb10s02
BibTeX
@article{euvimed2010,
title = {2. Executive summary},
journal = {New South Wales Public Health Bulletin},
year = {2010},
volume = {21},
number = {1},
pages = {9--10},
doi = {10.1071/nb10s02}
}
EUVIMED