Online real-time learning strategies for data streams for Neurocomputing

The influence of data streams has been more and more apparent in real-life problems than it was decades ago and has resulted from rapid development of smart sensors, information technology, internet of things, etc. [1]. Large amount of data points are sampled in a rapid manner and result in a data explosion which demand online real-time strategies to prevent loss of accuracy, system's instability due to slow feedback and intractable computational complexity. Online real-time strategies are preferred over their batch counterpart because it allows a large amount of data streams to be handled with O(1) complexities while retaining reliable accuracy [2]. It addresses the given problem in the one-pass fashion which makes such approaches practical even under limited computational resources. This trait is a clear competitive advantage for data stream analytics and even for big data analytics against the offline approaches through distributed and/or parallel computations which usually necessitate computer clusters under laborious pre-setting requirements [3]. The online real-time strategies need to be combined with evolving and adaptive strategies in order to overcome gradual, abrupt and seasonal changes in data patterns of data streams which need to be addressed effectively. The absence of such mechanisms leads predictive accuracy to deteriorate significantly when observing concept change in data streams [4]. The synergy of evolving and adaptive trait in online real-time strategies results in a tradeoff between plasticity and stability which ideally overcomes catastrophic forgetting of previously valid knowledge. It has been a primary ingredient to the success of data stream mining.