Abstract
This paper introduces a concept of neural network specialization viatask-specific domain constraining, aimed at enhancing network performance ondata subspace in which the network operates. The study presents experiments ontraining specialists for image classification and object detection tasks. Theresults demonstrate that specialization can enhance a generalist's accuracyeven without additional data or changing training regimes: solely byconstraining class label space in which the network performs. Theoretical andexperimental analyses indicate that effective specialization requires modifyingtraditional fine-tuning methods and constraining data space to semanticallycoherent subsets. The specialist extraction phase before tuning the network isproposed for maximal performance gains. We also provide analysis of theevolution of the feature space during specialization. This study paves way tofuture research for developing more advanced dynamically configurable imageanalysis systems, where computations depend on the specific input.Additionally, the proposed methods can help improve system performance inscenarios where certain data domains should be excluded from consideration ofthe generalist network.