Open world intelligence through composition of nonparametrics

This proposal responds to the direct to the Snap-On Phase 2 SBIR by extending our novel non-parametric Bayesian framework---OTACON: (O)pen world intelligence (T)hrough (A)ugmentation and (Co)mposition of (N)on-parametrics---for novelty detection in open worlds.  We build on our advances in composition and inference of non-parametric Bayesian models, and insights from the mathematics of Optimal Transport.  Our non-parametric approach applies to all levels of the SAIL-ON novelty hierarchy and through composition, extends to joint inference across levels simultaneously, as well as both static and dynamic domains. Our proposed system will address the following key technical challenges: developing domain-independent technical approaches that can address detection, characterization, and accommodation of novelty from all novelty hierarchy levels; and identifying "snap-on" technologies that can be added to already existing agents in both action and perception-oriented domains. First and second wave AI fail to generalize to novel situations due to brittleness and extreme sensitivity to discrepancies between the training and test sets.  These methods are, therefore, not applicable to the kinds of novelty that one encounters in open-world domains.  Bayesian non-parametrics -- such as Dirichlet processes, etc.-- by design assume open worlds with potentially unbounded complexity, and therefore provide a solution. Under this SBIR, we  will focus on transitioning technologies and providing strong mathematical foundations and performance guarantees for open-world learning agents capable of detection, characterization, and accommodation to novelty in open-world domains.  We will additionally collaborate with government and SAIL-ON performers to develop theory of open-world novelties, participate in SAIL-ON PI meetings and program-wide evaluations, and maintain compatibility with DARPA SAIL-ON objectives.  We have already achieved the required baseline capabilities through participation in the SAIL-On program. We will show that we have:  Demonstrated and quantified ability to detect and accommodate unforeseen novelties resulting from changes in objects, agents, actions, relations, and interactions, using nonparametric techniques in at least 3 distinct test environments, with independently validated results; Identified approaches for characterizing unforeseen novelties;  Preliminary theoretical mathematical framework for open-world learning;  Tested, benchmarked, and documented system Architecture, API, and UI;  and  Demonstrated compatibility with DARPA SAIL-ON objectives.  Our domain independent, Bayesian nonparametric approach is therefore ideally positioned to advance the goals of the SNAP-ON program.