Multi-task Learning For Multi-label Classification And Time-series Forescating

Multi-task learning is an crucial branch in the traditional machine learning ap-proaches.This can transfer the implication or ad-hoc property of a source domain to the target domain by learning a quality representation.Many researchers focus on this domain to provide supplementary for related tasks,or transfer the related knowledge to other domains to solve the unannotated samples.In the realistic machine learning applications,there are several important tasks:fewer annotated samples,data imbalance problem.The target of multi-task learning is to learn a better representation which can transfer the important and related knowl-edge to other domains.n the visual field,the knowledge of target detection and deep speculation learning can assist the other party,and multi-task learning can also provide different perspectives for the auxiliary task.Multi-label classification is an important topic in traditional machine learning.Traditional methods often use the kernel norm to express the relationship between tags.This paper proposes a method for accelerating the traditional kernel norm(MLAPP),considering each tag as a task and considering the connection between tasks.The effect is comparable to the traditional kernel norm,but the speed is faster.The experiment proves that the method of accelerating the ker-nel norm proposed in this paper not only considers the connection between the tags,but also speeds up significantly.Multitasking can also achieve better results in time series detection tasks.Tradi-tional time series detection is divided into two kinds of ideas.One is to describe the trend of time series by adding special regularization items.But these regularizations often require prior knowledge,and at the same time have great limitations and can only depict a certain change.The other is through a random method,such as a Gaussian process.But these stochastic methods can only assume changes in a certain type of distribution and cannot cope with the diverse changes in real life.This paper proposes a method of substrate formation that ultimately allows these selected substrates to cope with diverse changes by selecting the appropriate substrate.Experiments show that the proposed method can cope with the strong diversity of sample size changes,and the effect of various natural time series detection tasks is significantly improved.