دانلود کتاب مبانی، معماری و نمونه‌سازی هوش مصنوعی انسانی‌شده: یک رویکرد ساختارگرایانه جدید

Humanized AI (HAI) is a phrase used to denote artificial social beings that are very close to humans in various displayed aspects. These may include learning ability, knowledge discovery, problem-solving, creativity, communicated emotion, self-awareness, and consciousness. In a sense, AI is the artificialization of humans, while HAI is a re- humanization of artificialized humans. HAI may be considered as human-level AI or artificial general intelligence, but it also emphasizes human-like features. HAI agents are machine-race human beings, not digital slaves. It is noticed that the term “Humanized AI” might be used in a different context or a narrow sense by different people. This book is about how to make HAI agents which come as the next AI wave. It covers philosophical discussions on central topics in HAI, the architectures of HAI, and prototypical HAI agents. The agents are virtually Zero-data based agents (Zda) and Languageindependent agents (Lia), having no-built-in natural languages. Both are capable of displaying great learning skills, self-awareness, consciousness, and emotional intelligence. Unlike narrow AI, where each agent can only have a particular skill, the elaboration tolerance of HAI enables a humanized agent to learn many different skills over time without restarting or erasing previous information. I started to work on HIA in 2003, some 20 years ago, when I first realized a simple known fact: a human baby has virtually a data-empty brain and no inherited natural languages, but can learn things like languages and math skills. I immediately prototyped a very simple version of HAI using Microsoft Visual Basic. As I expected, the agents can learn and create very simple “language” through interactions by means of reinforcement learning. The approach is what is now called the Constructivist Approach. However, shortly thereafter I faced a challenge: when complex concepts are involved, traditional reinforcement learning has too many paths to explore before the agent can identify an optimal or reasonable solution. I struggled to solve the problem for the next 15 years until I finally came up with the current synthesist’s approach in 2016, which combines the behaviorist’s efficiency and constructivist’s flexibility. Behaviorists believe actions are reflections of what goes on in the mind, adopting a goaldriven approach. Constructivists emphasize that knowledge cannot be a passive reflection of reality, but an active construction by the individual, from simple to complex. We take a synthetic approach by combining the two in our HAI architecture—a new constructivist approach. In this approach, instead of emphasizing the notion that an agent’s action is based on the maximization of some utility as Rationalists do, we adopt probabilistically a randomized maximization to better emulate human-thinking and behaviors. The HAI architecture has four main components. (1) A Recursive Network of Patterns, for dynamic knowledge representation, provides a self-inclusive structure necessary for self- awareness. (2) Attention Mechanisms allow an HAI agent to focus on a few important things for efficient learning and response. (3) Learning Mechanisms that feature hierarchical tokenization and recursive patternization allow progressive learning, from simple to complex. (4) Adaptive Reinforcement Response Mechanisms, mimicking free will, randomize potential responses according to the associated rewards or frequencies. A major feature, or objective, of HAI is that imitation, doing what other people do under various situations, makes HAI agents adhere to social norms without pre-specifying any of the norms. Our HAI approach is constructed on several general principles and laws, including the Similarity Principle, Parsimony Principle, Association Principles, First Principles, Principle of Factor Isolation, Law of Summative Effects, and Weber-Fechner Laws. The principles make all specific methods or mechanisms in the HAI architecture coherent. The Similarity Principle, used in learning and response mechanisms, is the key to dealing with novelties encountered by HAI agents in patternization or scientific discovery. This is because everything is changing, and no two things are identical. This principle asserts that two similar objects behave similarly, thus allowing us and agents to group similar things into categories and make patterns. The Similarity Principle deals with imitation and creativity under the same umbrella, i.e., in similarity replacement of the building blocks of the world with different degrees of similarity. The Association Principle guarantees the association between two things that are spatially or temporally close, and it serves as the backbone for learning. The Association Principle allows an agent to map natural language patterns to environmental eventpatterns. In this sense, HAI can be considered as a computer program that automatically generates computer programs. What appears to be a complex universe can actually be modeled by a recursion of a simple world governed by a small set of principles. Enlightened by First Principles and Connotation of understanding, we postulate that any complex concept or knowledge consists of a hierarchical recursion of elementary concepts and any high-level skill is a composition of sequential and recursion of elementary movements of body parts. These atomic concepts and movements, despite varying from individual to individual, are the building blocks of our HAI agents—they make it possible to build HAI using a small-data approach! The Weber-Fechner Laws serve the basis for building virtual sensory organs, the Parsimony Principle supports the adaptive reinforcement learning in response model, the Law of Summative Effects ensures that the sequence of short-term goal-driven actions will achieve a long-term goal, and the Principle of Factor Isolation guides experience-based learning (patternization) and cognitive learning (logical reasoning). Given the complexity of HAI, studying HAI without discussions of architecture and prototyping to show how it works will be somewhat empty and less convincing, while studying HAI architecture and prototyping without understanding the fundamental issues and human characteristics is a mindless approach that will not lead too far either. Thus, I have an ambitious goal: to cover vertically the whole spectrum from foundation, architecture, and prototyping (algorithm and pseudocode) in a concise book. Keep in mind that even if (or when) we make the first agent exactly like a human baby, it will take 10–20 years to teach him what he needs to thrive. Principles and methods of teaching HAI agents are also a critical topic for our discussion. The book has four parts and an appendix. Part I features philosophical discussions of some key aspects of HAI, such as consciousness, the connotation of understanding, roles of attention, imitation, analogy and creativity, and scientific principles. Part II reviews the AI waves, existing approaches to HAI, and the new synthesist’s approach. These two chapters constitute the foundation for the HAI Architecture in Part III, which elaborates the architectures of the new constructivist approach, including innate knowledge, knowledge dynamic representation, learning mechanisms, response mechanisms, and effective teaching. Following the architecture blueprint, Part IV discusses prototyping the animated agents, Zda and Lia, with examples to show how language can be learned from scratch, how agents can learn playing board games starting from learning the rules, and many more. To make the book stand alone, The Appendix provides a concise review of narrow AI with emphasis on the ideas behind each method. This book is intended for anyone who is interested in HAI, including college professors, researchers, and students in Computer and Information Science, Computer Engineering, Data Science, Philosophy, Psychology, Education, Economics, and Political Science. The first two parts are geared toward the general audience, while Parts III and IV are more suitable for those who want to know exactly how HAI can be built to have human characteristics such as understanding natural language, self-awareness, consciousness, imitation, creativity, discovery, reasoning, goal-setting, and ethics. If you finish all four parts, you are expected to be able to use the computer language you are familiar with to translate the algorithms and pseudocode in Part IV into actionable agents, and to start to “raise” and teach your baby agents.