Scalable Intelligent Systems by 2025

Abstract

Reusable Scalable Intelligent Systems (RSIS) have the following characteristics: • Interactively acquire information from video, Web pages, hologlasses (electronic glasses with holographic-like overlays), online data bases, sensors, articles, human speech and gestures, etc. • Real-time integration of massive pervasively inconsistent information • Self-informative in the sense of knowing its own goals, plans, history, provenance of its information and having some information about its own strengths and weaknesses. • Close human interaction using hologlasses (electronic glasses with holographic-like overlays) for secure mobile communication. Computers alone cannot implement the above capabilities. • No closed-form algorithmic solution is possible to implement the above capabilities • Reusability so that advances in one area can readily be used elsewhere without having to start over from scratch • Scalability in all important dimensions meaning that there are no hard barriers to continual improvement in the above areas, i.e., system performance continually significantly improves. A large project (analogous to Manhattan and Apollo) is required to meet strategic challenges for Intelligent Systems and S5G (Secure 5G). Inconsistent and incoherence are the natural result of the enormous amount of information necessary for the operation of a Reusable Scalable Intelligent System. For any empirical proposition, there typically can be inconsistent information. Consequently, possibly inconsistent information is the default. Hologlasses are electronic glasses that provide holographic-like overlays, which will be used for just about everything including the following: driving (vehicle detection, traffic sign recognition, obstacle avoidance, navigation), warfare (navigation, reconnaissance, hazard avoidance, command and control), exercising ( navigation, timing, goals, repetition counting, advice), walking (obstacle avoidance, navigation, oncoming vehicle detection), reading (privacy, hyperlinks, synonyms and antonyms, annotations), cooking (menus, demonstrations, suggestions), shopping (price comparison, videos of product use, navigation), construction (placement, coordination, scheduling), maintenance (diagnosis, placement, verification), teamwork (shared presentations, agenda, meeting notes), entertainment (privacy, interactive movies, music, games), medical (medication management, diagnosis, pain management, physical therapy, dementia management, record keeping, meditation, visualizations), conversation (facial movements, eye tracking, pupil contraction and dilation), education (tutoring, demonstrating, collaborating) According to [Shamir, et. al 2019] “given any two classes C1 and C2, along with any point x∈C1, and our goal is to find some nearby y which is inside C2. Our ability to do so suggests that all the classes defined by neural networks are intertwined in a fractal-like way so that any point in any class is simultaneously close to all the boundaries with all the other classes.” The results show gradient classifier can be used against itself to find nearby inputs that are classified differently that one classified by a given input. [Shamir, et. al 2019] results suggests that gradient classifiers (aka "Deep Learning") could remain forever fragile in ways that cannot be remedied by larger training sets. Consequently, a gradient classifier can be useful for heuristic rule-of-thumb guidance but should not be relied upon as the exclusive basis of consequential judgments especially ones involving life and/or death. A gradient classifier (regardless of size) is not scalable to Reusable Scalable Intelligent Systems, e.g., for pain management and dementia management because data available for gradient training is very sparse since operation of a Reusable Scalable Intelligent System depends heavily on history of the case at hand, e.g., what worked previously, progression of ailments, recent interactions, etc. Also, multiplicity of potential actions (e.g. in pain management and dementia management) is very large because at any point interaction can take many different directions, e.g., social interaction, physical therapy, medication management, etc. A large gradient classifier is not modular. For example, a pain management gradient classifier cannot be incorporated into a dementia management gradient classifier because the pain management classifier does not provide sensory inputs that are usable by the dementia management gradient classifier. Furthermore, in order to incorporate additional knowledge into the dementia management gradient classifier, the knowledge must be illustrated in operation using sensory input to create input-output training data. Creating an input-output training data to incorporate additional general knowledge can be extremely difficult because of unknowable interactions with previous input-output training data including history dependence and multiplicity of potential actions. According to [Jordan 2018, Davis and Marcus 2019], overinvestment in gradient classifiers (aka “Deep Learning”) is crowding out necessary investment in other technologies for Intelligent Systems Multitudinous different kinds of classifiers need to be created and operated in an architecture that provides services for their training, optimizing, debugging, refactoring, testing, and monitoring.