Ensemble-based Human Communication Recognition

We propose a novel architecture for systems that target the recognition of human communication Distributed Ensembles. Distributed Ensembles results from the observation that in many different fields hard problems are handled by employing multiple computational entities that cooperate to solve a problem. Even though these solutions share this common trait, the goals in each field for employing multiple computational entities can be very different from each other, and can be as distinct as reducing error-rates in Automatic Speech Recognition, obtaining faster convergence in optimization problems, and achieving high performance within multimodal recognition systems through parallel algorithms. As a consequence of the variety of goals, the structure of the entities and the nature of cooperation are also varied and it is usually the case that solutions do not reap the full benefits that we see could potentially be obtained. While existing solutions emphasize different aspects, our observation is that these aspects are not mutually exclusive, but complementary. Three of these aspects seem particularly useful in a general sense: performance and scalability through multiprocessing; faster convergence by sharing partial results; error reduction by combination of hypotheses. We therefore propose a style that has potential for combining these three aspects, based on ensembles of distributed computational entities. To enable the use of the envisioned approach, we propose to develop an architectural infrastructure to offer efficient coordination services that are exposed as an API to facilitate use from a developer’s perspective. We explore issues surrounding Distributed Ensembles in the context of automatic recognition of human communication, that we show offer a particularly promising field of application. More specifically, we focus on American Sign Language (ASL) Recognition, a problem amenable to well researched natural language processing approaches.