Conditions Of The Linear Chain With The Airport Training Algorithm Optimization

The proposal of Conditional Random Fields (CRFs) is a milestone in the history of machine learning. CRFs integrated the advantages of several kinds of probabilistic models, and excluded the disadvantages of them. Because of the intrinsic advantages they possess, CRFs are widely used in machine learning field, especially in natural language processing field.Despite the advantages, CRFs also suffer from some drawbacks. One of the significant drawbacks is that the training process of a CRF model requires very high computing resources. First, the training process is memory-consuming. The memory requirement may exceed the available physical memory size, when training on relatively large tasks. Second, the computation of the training process itself is time-consuming, sometimes it needs days or even weeks to finish for large tasks. This is not only due to the complexity of the training algorithm, but also due to the inefficient use of the features of modern hardware. These make it difficult or even infeasible to use CRFs in large-scale data analysis tasks.The object of this thesis is to propose an efficient algorithm, to enhance the efficiency of the training process of CRFs, so that they can be applied to large tasks. We deal with this problem from three aspects:First, software prefetching techniques are utilized to hide cache miss latency. Second, we exploit modern CPU's features to process data in parallel. Third, when dealing with large data sets, we let HOCT instead of operating system to manage data swapping operations between memory and disk.Experiments show that the algorithm proposed by this thesis do improve the efficiency, especially when the task gets larger, the improvements gets stronger. This demonstrates that the algorithm proposed by this thesis can be applied to large training tasks.