Computer Vision News - August 2018

Method: The method is comprised of four major steps (presented in the next page) I. Task-Specific Modeling: training a specific network for a single visual task II. Transfer Modeling: computing the transfer learning function between each pair of tasks III. Normalization: normalizing the graph IV. Taxonomy Extraction: subgraph selection -- Boolean Integer Programming was used to identify the best source tasks and what sources to use for each target task. The figure in the next page illustrates the four steps of the method: at the top, the 26 visual tasks are represented as nodes of a graph, with edges representing the transfer learning affinity between pairs. Below -- we zoom in on what occurs within a single node (visual task network) during original training (step I) and transfer learning (step II). Computer Vision News Research 5 Research Computer Vision News