The Dawn of a New Decade with Digital Sentient Systems Powered by Post-Turing Computing Models, Strong AI & Self-Managing Edge Clouds

"The ongoing success of applied Artificial Intelligence and of cognitive simulation seems assured. However, strong AI, which aims to duplicate human intellectual abilities, remains controversial. The reputation of this area of research has been damaged over the years by exaggerated claims of success that have appeared both in the popular media and in the professional journals. At the present time, even an embodied system displaying the overall intelligence of a cockroach is proving elusive, let alone a system rivalling a human being."

B. J. Copeland, May 2000, AlanTuring.net, Reference Articles on Turing (What is Artificial Intelligence? Is Strong AI Possible?)

https://www.alanturing.net/turing_archive/pages/reference%20articles/what_is_AI/What%20is%20AI13.html

Introduction:

Last year, I posted four trends that are contributing to reducing IT complexity, improving data privacy and integrating human-machine interaction:

1.  Potential virtual machine (VM) layer elimination,
2. Potential Legacy and virtual machine network overlays,
3. Use of crypto-security and digital asset management without the complexity of crypto-puzzles and current block-chain implementations, and
4. Evolution of machine consciousness and culture and their integration with human consciousness and culture.

This year I attended the International Society for the Study of Information 2019 Summit at the University of California, Berkeley and listened to a few interesting talks on the progress of AI, current limitations and few questions about its future. Deep learning, has delivered a variety of practical uses in the past decade by revolutionizing customer experience, machine translation, language recognition, autonomous vehicles, computer vision, text generation, speech understanding, and a multitude of other AI applications. Deep learning models do not require algorithms to specify what to do with the data. Extraordinary amount of data we as humans, collect and consume — is fed to deep learning models. An artificial neural network takes some input data, and transforms this input data by calculating a weighted sum over the inputs and applies a non-linear function to this transformation to calculate an intermediate state. The three steps above constitute what is known as a layer, and the transformative function is often referred to as a unit. The intermediate states—often termed features—are used as the input into another layer. Through repetition of these steps, the artificial neural network learns multiple layers of non-linear features, which it then combines in a final layer to create a prediction.

A talk by Melanie Mitchell based on her book “Artificial Intelligence – A Guide for Thinking Humans” summarized for me the issues with Deep Learning and the current state of AI. The neural network learns by generating an error signal that measures the difference between the predictions of the network and the desired values and then using this error signal to change the weights (or parameters) so that predictions get more accurate. Therein lies the limitation of Deep Learning. While we gain insights about hidden correlations, extract features and distinguish categories, we lack transparency of reasoning behind these conclusions. Most importantly there is the absence of common sense. Deep learning models might be the best at perceiving patterns. Yet they cannot comprehend what the patterns mean, and lack the ability to model their behaviors and reason about them.

Based on our knowledge of how natural intelligence works, we can surmise that the following key elements of human mind, which leverage the brain and the body at cellular level, are missing in current state of the art A.I.:

1. Time Dependence & History of Events: In Nature, systems are continuously evolving and interacting with each other. Sentient systems (with the capacity to feel, perceive or experience) evolve using a non-Markovian process, where the conditional probability of a future state depends on not only the present state but also on its prior state history. Digital systems, to evolve to be sentient and mimic human intelligence, must include time dependence and history in their process dynamics.
2. Knowledge Composition and Transfer Learning: The main outcome of this ability is to understand and consequently predict behaviors by a succession of causal deductions supplementing correlated inductions.
3. Exploration vs. Exploitation dilemma: Creativity and expertise are the consequences of our ability to swap from the comfort zone to unchartered territories and it’s a direct and key usage of our transfer learning skill. Analogies and Translations are powerful tools of creativity using knowledge in a domain and applying it in another.
4. Hierarchical structures: As proved by G?del, an object can only be described (and managed) by an object of a higher class. A key principle of how cells are working by exchanging proteins whose numbers, functions, and messages are supervised by DNA at cell level or group (higher) level.

True intelligence involves generalizations from observations, creating models, deriving new insights from the models through reasoning and changing the models based on experience or the availability of new knowledge. In addition, human intelligence also creates history and uses past experience in making the decision. Human intelligence has leveraged computing (information processing), communication (information networking), cognition (ability to perceive and learn) to develop consciousness (awareness of self and the environment both at an individual level and at a group level) and culture (developing best practices and adopting them at a global level while balancing the individual and group requirements). Artificial intelligence in order to mimic human intelligence must incorporate the five C's in its framework.

Listening to various discussions brought back the memories of similar discussions of Expert systems in the late 1980’s when I was in the Bell System designing expert system tools and applications. The following abstract from that period summarizes those discussions.

John C. KUNZ, Marilyn J. STELZNER, Michael D. WILLIAMS, From Classic Expert Systems to Models: Introduction to a Methodology for Building Model-Based Systems, Editor(s): Giovanni GUIDA, Carlo TASSO, Studies in Computer Science and Artificial Intelligence, North-Holland, Volume 5, 1989, Pages 87-110. (https://www.sciencedirect.com/science/article/pii/B9780444873217500090 )

“Classical expert systems have been built by representing empirical associations as described by experienced human experts. This paper presents examples of knowledge systems which are model-based: they represent and manipulate descriptions of the structure of a domain and the principles which characterize its behavior. Model-based systems have been developed as a way both to exploit existing knowledge of the principles of a domain and to extend the limited flexibility which is often found in classical expert systems. The model-based approach emphasizes representation of the structure and function of the modeled system and reasoning with respect to that structure and function. We discuss the methodology which has been used to develop such model-based systems and a number of important design issues. We particularly emphasize the model-based representation and reasoning and the interactive interfaces required to support these models. Examples are first presented of two different models, of different kinds of reasoning and interfaces which were used with those models. Following discussion of models, model-based reasoning and interactive interfaces, there is a discussion of the way that model-based reasoning and its supportive interfaces extend traditional expert systems and the engineering discipline of modeling. “

Key aspect of transitioning to model-based reasoning is to use the domain knowledge in the form of functions and structure depicting the behaviors based on events. Digital systems, to evolve to be sentient and mimic human intelligence as closely as possible, must include time dependence and history in their process models and dynamics. I argue that introducing the results of the theory of Oracles and the non-Markovian computing model, we mimic the natural intelligence more closely in the digital universe. All advances in physics are preceded by associated mathematical theories. Recent advances in superrecursive algorithms, the theory of Oracles, Knowledge Structures, Named Sets and the theory of Structural Machines point to a new advance in the physics of information science.

A recent review by Prof. Burgin details the theories and their application pointing to a new paradigm.

Burgin, M., Eberbach, E., & Mikkilineni, R. (2019). Cloud Computing and Cloud Automata as A New Paradigm for Computation. Computer Reviews Journal, 4, 113-134. Retrieved from https://purkh.com/index.php/tocomp/article/view/459

In this post, I summarize two new computing models (which the Church-Turing thesis proponents and the Silicon Valley pundits have not caught on to yet and exploit) and discuss applications of the new theories going beyond our parent’s or contemporary computer science. If I were a graduate student, this is where I would turn my attention to.

Post Turing Computing Models:

There are two computing models derived from the theories of Superrecursive algorithms, theory of Oracles, theory of knowledge structures, theory of named sets and the theory of Structural Machines alluded to above.

Structural Machines as Self-Managing Software Systems:

Church-Turing thesis boundaries are challenged when rapid non-deterministic fluctuations drive the demand for resource readjustment in real-time without interrupting the service transactions in progress. The information processing structures must become autonomous and predictive by extending their cognitive apparatuses to include themselves and their behavior (non-functional requirements) along with the information processing tasks at hand (functional requirements). While there was a lot of resistance to accept the call for new computing models going beyond Church-Turing thesis originally pointed out by Peter Wegner as long ago as 1997, it is now well established that new computing models based on the theory of Oracles are more efficient, resilient and scalable than machines based on just stored program implementation of Turing Machines (for a discussion of the theory and implementation, see Burgin, M. and Mikkilineni, R. (2018) Cloud computing based on agent technology, super -recursive algorithms, and DNA, Int. J. Grid and Utility Computing, v. 9, No. 2, pp.193–204). Self-management of information processing structures in the face of non-deterministic fluctuations are implemented using hierarchical network of cognizing agents which configure, monitor and reconfigure as required to meet the fluctuations both in the functional and non-functional behaviors.

Model Based Strong AI with Knowledge Structures, Deep Reasoning and Deep Memory Augmenting Deep Learning:

New hierarchical computing structures with cognizing agents go beyond neural networks to provide models of the observations, abstractions and generalizations from experience and create time dependent evolution and history to provide reasoning and predictive behaviors. The models comprising of deep knowledge are designed to capture not only classification of objects, their attributes and relationships but also behaviors associated with them. These behaviors are captured as generalizations from history and observations. At any point of time, any new event triggers an evolution of the current state to a future state based on not only the current state but also dependent on its past history.

This non-Markovian behavior gives rise to a new level of intelligence that goes beyond mere computing, communication and cognition alone support. In order to model this level of intelligence, we propose a superrecursive neural network, an ontology-based model of the domain of interest created from various pieces of knowledge (observations, experience, science, common sense etc.) and memory that captures time and history of various instances populating the model.

Armed with these computing models, we can now implement a new class of sentient, resilient and intelligent systems that manage themselves while ensuring the intent of the information processing structure is not interrupted.

Self-Managing Federated Edge Clouds

A new device by Platina Systems provides a vehicle to implement the controllers and processors and create a structural machine that executes both functional and non-functional requirements. When a resource fails to deliver the required fuel to the execution of functional requirements, the controller automatically reconfigures the structural machine to provide the service without interruption. Figure 1 shows Platina edge computing cluster.

Figure 1: Hierarchical network of cognizing agents providing sentient, resilient and intelligent structural machines

Platina’s high-performance edge cluster architecture provides a foundation for implementing “Structural Machine” with a hierarchy of controllers and a processing cluster. The processing cluster:

1. Provides bare-metal and low-latency computing using commodity off-the-shelf computing hardware cluster connected at 100 GbE over fiber;
2. Eliminates complexity of legacy networking stacks by replacing merchant switch ASIC SDKs with a streamlined user space driver layer and using only L3 routing with leaf-spine topologies;
3. Integrates low-latency storage (SSDs with RDMA and NVMeoF);

Platina Controller provides the control of the distributed clusters that are directly connected through L3 Network and integrates it with the Kubernetes container orchestration. This brings two benefits:

1. Allows zero-touch provisioning and run-time QoS assurance of services in each cluster based on service and resource blueprints, and
2. Seamlessly connects to state of the art L2/L3 networks or provide end-to-end L3 network services bypassing L2 network Operation and management complexity. This allows a simpler edge network of clusters without the need for L2 services, in essence eliminating the LAN. It also provides a path with container orchestration to move workloads from L2 infrastructure to the new L3-only infrastructure.

In addition, the controller provides a unique data access mechanism at in-memory speed. First, using direct attached non-volatile memory, data access can be provided at in-memory speed. Second, it supports dedicated storage ports with dedicated storage name space to create data access at almost in-memory speed.

Named Microservice Network Provisioning, Operation and QoS Assurance at Run-time are provided using the controller to provision IaaS with following features:

1. Auto-discovery of environment and infrastructure resources,
2. Auto-configuring of resources, and
3. Microservice network deployment and monitoring

A hierarchy of cognizing security agents collaborate in configuring, monitoring and managing downstream security processes while sharing global knowledge which, allows the ability to influence one component security with the knowledge of what is happening in another component in a distributed system.

In essence, the hierarchy of cognizing agents leverage container orchestration, managed L3 Networking with bandwidth slicing and shared storage at in-memory speed to deploy, monitor and assure run-time QoS for workloads.

Currently, this architecture is being used in deploying a diverse set of application stacks (web services, federated AI and trust-but-verify business processes for implementing business governance, risk and compliance (GRC)) at Golden Gate university (called project DeepEdge) in collaboration with MetricStream, a leader in GRC) using containerized web stacks and deep learning stacks managed by cognizing agents.

The DeepEdge project is aimed at defining the next generation GRC architectures which must address external threats and vulnerabilities by extending the information processing technologies and processes to include behavioral and interaction models of the both internal and external actors. GRC system of the future must be able to model, analyze various events and predict high risk consequences of important interactions among key actors.

Today, Governance, Risk and Compliance (GRC) processes in the enterprise are siloed, sporadic, segregated and are implemented as back-office after-thought often, requiring labor-intensive supervision. While business process management and its automation have improved resiliency, agility and efficiency, GRC is not integrated in the mainstream. Current practices tend to focus on analysis after the fact to audit compliance. They are not designed to prevent many consequences that result from intentional actions of various actors such as fraud, illegal transactions, breach of contracts etc. In addition, business decisions made in optimizing one business process may have consequences that increase overall risk for the enterprise. Current information systems do not facilitate detection and flagging of such events because they are not readily available or buried in few expert’s knowledge or labor-intensive. Their interactions depend on incentives, motivations and cost benefit advantages that would affect them. While it is possible to manage the risks within the BPM processes, albeit labor intensive, it is far more difficult to manage the risks outside which far outweigh the internal risk management effort.

The new architecture defines knowledge structures and populates them using Deep Learning and follows their evolution using Cognizing Agents with global knowledge and local control. Figure 2 shows the DeepEdge project.

Figure 2: The DeepEdge project.

It is hoped that pushing the trust-but-verify processes to the edge allows local control of data and provides security and the cognizing agents provide global knowledge sharing that allows end-to-end visibility and auditing control using real-time insights.

In summary, theory and implemeations are converging to provide post-Turing information processing structures that :

1. Are self-managing and tolerant to fluctuations in the demand for and availability of computing resources using cognizing agents implementing structural machines,
2. Augment Deep Learning with model based Knowledge Structures, Deep Memory, and non-Markovian Deep Reasoning with cognizing agents and
3. Enable sentient, intelligent, resilient, distributed, federated and self-managing business processes with local control and global knowledge sharing in real-time using cognizing agents.

Finally, in spite of the resistance from "deep-state computer scientists," perhaps Late Professor Peter Wegner's call for new computing models will be answered in the coming decade.

Late Prof. Peter Wegner was a Keynote Speaker and presented his thoughts about post-Turing Computing Models and Chaired a panel discussion at the workshop on Architecting Self-Managing Distributed Systems in the 9th European Conference on Software Architecture (ECSA 2015)