Emerging Soft Computing Methodologies in Deep Learning and Applications
摘要截稿:
全文截稿: 2019-09-30
影响因子: 5.472
期刊难度:
CCF分类: 无
中科院JCR分区:
• 大类 : 计算机科学 - 1区
• 小类 : 计算机:人工智能 - 2区
• 小类 : 计算机:跨学科应用 - 1区
Overview
Machine learning is to design and analyze algorithms that allow computers to "learn" automatically, and allows machines to establish rules from automatically analyzing data and using them to predict unknown data. Traditional machine learning approach is difficult to meet the needs of Internet of Things (IoT) only through its outdated process starting from problem definition, appropriate information collection, and ending with model development and results verification. But however, recent scenario has dramatically changed due to the development of artificial intelligence (AI) and high-speed computing performance. Therefore, deep learning is a good example that breaks the limits of machine learning through feature engineering and gives astonishingly superior performance. It makes a number of extremely complex applications possible.
Machine learning has been applied to solve complex problems in human society for years, and the success of machine learning is because of the support of computing capabilities as well as the sensing technology. An evolution of artificial intelligence and soft computing approaches will soon cause considerable impacts to the field. Search engines, image recognition, biometrics, speech and handwriting recognition, natural language processing, and even medical diagnostics and financial credit ratings are all common examples. It is clear that many challenges will be brought to publics as the artificial intelligence infiltrates into our world, and more specifically, our lives.
Deep learning has been more mature in the field of supervised learning, but other areas of machine learning have just started, especially for the areas of unsupervised learning and reinforcement learning with soft computing methodologies. Deep learning is a class ofmachine learningalgorithmsthat:
use a cascade of multiple layers ofnonlinear processingunits forfeature extractionand transformation. Each successive layer uses the output from the previous layer as input.
learn insupervisedand/orunsupervisedmanners.
learn multiple levels of representations that correspond to different levels of abstraction; the levels form a hierarchy of concepts.
Due to the cascaded structure and the abstraction level of multiple representations, Deep Learning has very good performance in speech recognition and image recognition, especially when one aims to have different levels of resolution representations in signals and images with gaining automated features extracted from these. Two common models, Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN), are widely used architectures in the context of deep learning. In addition to the fact that most "deep learning" technologies are built on the concept of supervised learning to construct a set of classifiers to recognize things entering an information system, "soft computing and metaheuristic algorithms" are built on the concept of unsupervised learning to find out good solutions from a solution space, which can be regarded as an infinite space. The algorithms of these two research domains are the two promising technologies of AI that has been widely and successfully used in solving many complex and large-scale problems.
However, applying deep learning to solve problems will encounter some challenges. In order to have good performance, deep learning algorithms require a large and diverse range of data, and a large number of parameters need to be tuned. Furthermore, well-trained deep learning model tend to have overfitting problems, and not easily applied in other areas. In addition, the training process of deep learning is still a black box, and researchers have a hard time understanding how they learning and how they deduce conclusions. Therefore, in order to boost performance and transparency of deep learning models and to bring them actually to a level of high practical usage in real-world applications and facilities, this special issue places a special attention i.) on the (complexity) reduction of parameters with soft computing methodologies in deep-learning models, ii.) an enhanced interpretation and reasoning methods with soft computing methodologies for explaining hidden components in deep learning models as well as for gaining a better understanding of the outputs of deep learning models (=> increasing acceptability for company experts and users) and iii) on methods for incrementally self-adapting and evolving soft computing methodologies for deep learning models, where not only weight parameters may be recursively updated, but also internal structures may be evolved and pruned on the fly based on current changes and drift intensity present in the system. Furthermore, new deep learning methods in combination with renowned, widely-used architectures, but also developed for soft computing and artificial intelligence environments where it has been not considered so far (e.g., deep learning SVMs or deep learning bio-inspired systems are hardly existing) are also warmly welcomed. There are new emerging applications and new deep learning developments of established applications of soft computing methodologies and architectures, with specific emphasis in the fields of big data, internet of things, social media data mining, web applications.
Topics
Original contributions are solicited from, but are not limited, the following topics of interest:
Methodologies,and Techniques(but not necessarily restr. to):
New methods for Soft Computing in combination with Deep Learning
New learning methods with Soft Computing concepts for established deep learning architectures and structure
Faster and more robust Soft Computing methods for learning of deep models
Complexity Reduction with Soft Computing methods and Transformation of Deep Learning Models
Evolutionary and Soft Computing-based optimization and tuning of deep learning models
Evolving and Soft Computing techniques for deep learning systems (expanding and pruning layers, components etc. on the fly)
Metaheuristics aspects and Soft Computing algorithms in deep learning for improved convergence
Hybrid learning schemes with Soft Computing (deterministic with heuristics-based, memetic)
Interpretability Aspects with Soft Computing for a better Understanding of Deep Learning Models
Soft Computing Methods for non-established deep learning models (deep SVMs, deep fuzzy models, deep clustering techniques, ...)
Real-World Applicationsof deep learning techniques such as (but not necessarily restricted to):
Cloud and Fog Computing in AI
Big Data Analysis
Context-Awareness and Intelligent Environment Application
Financial Engineering and Time Series Forecasting and Analysis