| Home | Table of Contents |
General Intelligence :
version 2.0
In the following I'll try to establish some basic features of general intelligence. Let's see if everyone agree with them?
For a detailed theory (work in progress) refer to the table of contents.
Building a general intelligent agent can be approached just like any other engineering problem: by listing its requirements and then designing an architecture that can simultaneously satisfy all of them.
The basic requirements of G0 is as follows (summarized from ArtificiaI Intelligence: a Modern Approach 2nd edition by Russell & Norvig 2003):
In particular, G0 should not possess emotions so as to avoid ethical issues and to simplify the system.
Checkpoint 1: Do you think any other requirements are being neglected? Remember we're trying to build a minimalist general intelligence with only essential components.
Read here for a detailed requirements list.
My claim is that this architecture can satisfy requirements #1-5.
I'll explain these components one by one. Please be patient...
The cognitive engine consists of:
For the past 2 years I and my group have worked on this part:
Sensory experience is like a movie so it is represented as a "video cube":
Given a camera image or a video sequence, the Recognizer will be able to output something like "there is a house with a red roof in a night background and the moon" or "A hits B with his fist and then B kicks A with his leg..." etc. We have crudely figured out how to do this part (low-level sensory recognition).
The Recognizer is organized hierarchically, so low-level descriptions can be further abstracted. For example "A hits B and B hits A" etc can be recognized as "2 people are fighting". I think this - general pattern recognition - is the hardest part of general intelligence.
Central to the G0 framework is the notion that hierarchical pattern recognition can achieve compression. For example: the string ...ABCDDABCABC... can be compressed as ...XDDXX... if ABC is recognized as X. This is known as dictionary compression. My idea is to expand this idea to hierarchical and lossy compression.
A general intelligent agent needs to be aware of its own mental states. There are 2 types of self-awareness: affective ("I feel that...") and epistemic ("I believe that...").
Without self-awareness some mental concepts cannot be understood properly. For example "I think you're crazy" or "John lied to Mary" (with the intent of deceiving her).
The Emoter module provides a way to hypothetically model human emotions, which is not the same as experiencing emotions subjectively. Epistemic awareness is implemented by using autoepistemic logic (not shown here).
The Recognizer dumps descriptions into an internal representation, which is stored in several memory systems:
Concepts in the Recognizer (and in Generic Memory) are unlabeled, ie they have no names. The job of NLP is basically to label them, but the labeling may change dynamically (such as when learning a new word).
Queries are answered by probabilistic/fuzzy pattern matching and logical inference using memories as the knowledgebase.
The pattern matching algorithm may be a bit complicated because it has to handle polymorphisms, for example "the boy hits the girl" and "the girl is hit by the boy" actually refer to the same thing.
G0 will learn human values from a third-person perspective. For example "humans think this picture is beautiful" or "John likes Mary" etc. Then such knowledge can be used to solve real-life problems via the planner.
OK - this is the general framework. I want to establish consensus about the framework and then make sure everyone understands it. Please ask questions or express your ideas if they're different...
You may provide some instant feedback with this form:
| Home | Table of Contents |
6/Oct, 23/Aug, 16/Apr/2006 (C) GIRG [ Notice about intellectual property on this page ]