PROJECT TITLE :

Cold-Start Active Sampling Via γ-Tube

ABSTRACT:

Active learning, also known as AL, is a method that involves querying labels from a collection of unlabeled data in order to improve generalization performance for a given classification hypothesis. Generally speaking, an informative, representative, or diverse evaluation policy is utilized in order to evaluate the sampling process. However, because the policy can only begin operating after it has been provided with an initial labeled set, the performance of the policy may degenerate under a cold-start hypothesis. First, we will demonstrate in this article that the typical AL sampling can be equivalently formulated as geometric sampling over minimum enclosing balls. The MEB in this article refers to a conceptual geometry that is overlaid on the cluster in the analysis of generalization. It is related to the hard-margin support vector data description (MEBs) of clusters in the SVM community. After that, we take one MEB that covers a cluster and divide it into two parts, the first of which is a -tube, and the second of which is a -ball, following the structure of a -tube that is used in geometric clustering. Our theoretical insight reveals that -tube is able to effectively measure the disagreement of hypotheses in original space over MEB and sampling space over -ball. This is accomplished through the estimation of the error disagreement that occurs between sampling in MEB and -ball. As a means of refining our understanding, we have provided a generalization analysis, and the findings demonstrate that sampling in a -tube can produce a higher probability bound while still achieving nearly zero generalization error. Following these analyses, we finally present a tube AL (TAL) algorithm in order to combat the cold-start sampling issue. This algorithm is based on the informative sampling policy of AL over -tube. As a consequence of this, the dependency that previously existed between the process of querying and the evaluation policy of active sampling can now be eliminated. The results of the experiments show that by employing the -tube structure to manage cold-start sampling, TAL is able to achieve a higher level of performance than the standard AL evaluation baselines. This is accomplished through the presentation of substantial accuracy improvements. The recognition of image edges extends the scope of our theoretical results.