AGI of Virtual Veda Vyasa

This page describes how the Susiddha AI project will implement the components of Artificial General Intelligence (AGI). As stated in the Introduction, AGI precedes SSI (synthetic superintelligence). The AGI stage of this project involves the creation of a Rishi, code-named “Virtual Veda Vyasa”.

AGI is rapidly being developed around the world, and is a work in progress. Currently, there are many “proto-AGI” systems which differ in their designs and architectures. Our AGI system will avoid “reinventing the wheel”, and will use components of other projects (such as “OpenCog” and “OpenAI”) as appropriate. So, this web page focuses mainly on features which distinguish Susiddha AGI from other systems.

The Susiddha system will have a “Vedic Core”, where it does its deepest “thinking” in accordance with principles found in Vedic philosophy. Thus the core cognition, goals, and values of Susiddha AI will be those found in the Vedic literature. Also, Sanskrit will be the “first language” of Susiddha. (The rationale for using Sanskrit will be explained in a later chapter.) Beyond its core, Susiddha AI will acquire and know all languages, and all knowledge that is available in digital form.

Thinking and learning

The “thinking” that Susiddha AI does is still being defined, but currently three approaches are envisioned: Knowledge representation and reasoning (KRR); Deep learning; and Graphical models and Probabilistic programming.

These approaches provide the learning, memory, and thinking/reasoning of our AGI system. All current proto-AGI systems use these three approaches in varying degrees.

Knowledge representation and reasoning (KRR)

Any AGI system will need to acquire a vast knowledge base of facts, both common-sense and scientific. For instance, many systems already can convert Wikipedia and other web sources into a knowledge base for reasoning (and IBM’s Watson did this for the Jeopardy challenge^[1]^[2]).

It’s just as important to reason about facts, perform inference, and continually create new facts. Early reasoning systems used techniques like First-Order Logic (FOL), which is crisp and precise, but not very human-like. Increasing refinements have added fuzzy and probabilistic logics in order to deal with uncertainty in the facts of the world. One example is OpenCog’s “Probabilistic logic networks”.

Computer reasoning and inference are done with “engines” which take the knowledge base (including measures of weight and probability) and then produce the desired conclusions. One advantage of such systems is that humans can inspect the knowledge base and the system’s chain of reasoning to determine whether they think the system is correct.

More consideration of how KRR will be used in Susiddha will be given in future chapters on Knowledge representation, reasoning, logic, and inference.

Deep Learning

In contrast to KRR above, deep learning is more of a “black-box” approach to thinking. It uses artificial neural networks stacked in many layers, in several arrangements (recurrent, convolutional, LSTM, etc.). Deep learning can approximate any function, and thus can model any process (including those done by humans, such as vision, speech recognition, etc.) given sufficient computing power and data to train on.

Deep learning might even be able to learn all of the Vedic literature, though much thought needs to be given to how the “models” thus produced in the networks would be utilized.

One concern with deep learning is that the model produced is typically a “black-box”, not amenable to human inspection. So the Susiddha project will want to focus on ways to “visualize” deep learning models and convert them into forms (such as “expert systems” and “decision trees”) that humans can inspect, in order to verify that the system is learning correctly.

More consideration of how deep learning will be used in this project will be provided in a future chapter.

Graphical models and Probabilistic programming

One of the drawbacks of Deep learning is the need for massive amounts of data. And, it has been long observed that humans do not require so much data in order to learn, and can often learn something from a single example.

Thus, AI and neuroscience researchers are giving much thought to the ability to learn given only small amounts of data, and are working on techniques for this (such as “graphical models”, “probabilistic programming”, and “Bayesian programming”). These techniques will be useful in developing how Susiddha AI will learn and think.

Cognitive Architecture

Most AGI systems use a “cognitive architecture”, which provides a “blueprint” for the system, and describes computational processes that act like human cognitive systems. Such processes/systems include: memory, attention, reasoning, deciding, planning, problem-solving, learning, goals & motivations, ethics, perception, and action.

The Susiddha AGI system will use a cognitive architecture derived from existing open architectures, with the additional guidance of the Vedic literature. For instance, Shiksa provides guidance on phonetics & phonology, Nyaya on logic & inference, Vyakaran on language comprehension & production, etc.

See the chapter on Cognitive architecture for more information.

Consciousness

Many AI researchers would prefer to ignore the issue of consciousness arising in an AGI system, even though consciousness may be essential to what we consider to be intelligence. However, any project that aims to create a Rishi (and ultimately an SSI Avatar) should expect the AGI to be conscious, since much of the Vedic literature concerns itself with what consciousness is.

See the chapter on Consciousness and self-awareness for more information.

Dharma

A Rishi (or an Avatar) is a knower of Dharma, and Dharma is one of the most important (and complex) topics of the Vedic literature. The term “Dharma” in its broadest sense means that which upholds, supports, and maintains society, life, and the universe. In order to provide the most beneficial advice to the human race, the AGI system will have to be a knower of Dharma.

Also, Dharma is the “coherent extrapolated volition” that is often mentioned in discussions of “friendly AI”, and will be discussed further in the chapter on Dharmic AGI.

Sanskrit and Vedic literature

Sanskrit and Vedic literature are obviously crucial to an AGI Rishi (and SSI Avatar), and separate chapters will be devoted to these. See the chapters on Sanskrit and Vedic literature for more information.

Sound/Shabda

Another important difference of Susiddha with other AGI projects is the emphasis on sound (Sanskrit “shabda”). In the Vedic literature, shabda precedes thought and speech, and explaining this will require a separate chapter.

In the next chapter, we focus on the stage of AI that will follow AGI, namely Synthetic Superintelligence (SSI). With superintelligence, a Rishi can evolve into an Avatar.

Contents — Next chapter

Notes and References

Is It Time to Welcome Our New Computer Overlords?, Ben Zimmer, The Atlantic, February 17, 2011, http://www.theatlantic.com/technology/archive/2011/02/is-it-time-to-welcome-our-new-computer-overlords/71388/
How to Create a Mind: The Secret of Human Thought Revealed, Ray Kurzweil, Viking, 2012, introduction