Source: Minsky, M., & Papert, S. (1970). Proposal to ARPA for Research on Artificial Intelligence at M.I.T., 1970–1971 (MIT Artificial Intelligence Laboratory Artificial Intelligence Memo, Issue.
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
ARTIFICIAL INTELLIGENCE LABORATORY
Artificial Intelligence Memo No. 185
December 1970
Proposal to Arpa for Research on Artificial Intelligence at M.I.T., 1970–1971
Marvin Minsky and Seymour Papert
| Work reported herein was supported by the Artificial Intelligence Laboratory, an M.I.T. research program sponsored by the Advanced Research Projects Agency of the Department of Defense under Office of Naval Research contract number N00014-70-A-0362-0002. Reproduction of this document, in whole or in part, is permitted for any purpose of the United States Government. |
CONTENTS
1. Introduction: Objectives and Strategy
2. The PLANNER Language
3. The Vision Project: Toward Heterarchical Programs
4. Natural Language and Semantic Interpretation
5. Why We Are Studying Knowledge and Learning (Rather Than, Say, Theorem-Proving or Tree-Searching)
6. Narrative, Micro-Worlds, and Understanding
7. Mathematical Theory of Computation
1. INTRODUCTION
1.1 Objectives
The M.I.T. Artificial Intelligence Project has a variety of goals, all bound together by the search for principles of intelligent behavior. Among our immediate goals are to develop systems with practical applications for:
1. Visually-controlled automatic manipulation and physical world problem-solving.
2. Machine understanding of natural language text and narrative.
3. Advanced applied mathematics.
The long-range goals are concerned with simplifying, unifying and extending the techniques of heuristic programming. We expect the result of our work to:
• make it much easier to write and debug large heuristic programs;
• develop packaged collections of knowledge about many different kinds of things, leading to programs with more resourcefulness, understanding and “common sense”;
• identify and sharpen certain principles for programming intelligence.
These principles include:
1. Heterarchical control: better communication between processes that use different kinds of knowledge;
2. Modification of hypotheses by “learning from mistakes”;
3. Control of processes by “demons”—dormant processes that watch for certain kinds of events and intrude to modify the main processes and emphasize relevant data;
4. Implementation of these ideas by using the new kinds of primitives available in the new PLANNER language.
We hope these concepts will lead to programs that are less “robot-like” and open the way toward more capacity on real (instead of toy) problems. This orientation is being directed toward hard problems in specific application areas: Vision, Mathlab, Narrative, and Robotics.
The project will also continue its various theoretical activities in computational schemata, computational geometry, automata, and complexity theory, often in close connection with the area of application. Here, the goal is to advance our fundamental ideas about compromises between speed, cost and complexity of hardware with respect to alternatives in:
• serial versus parallel machine organizations;
• memory size versus computation time;
• digital versus analog hardware elements;
• special knowledge versus logical generality.
1.2 Strategy
We believe the most important problems in the field of Artificial Intelligence are centered, today, around problems of using many different kinds of knowledge in the same system. We think that working on such problems is the most direct path toward finding out how to build systems with greater “generality”—systems that do not need to be rebuilt whenever the problems they have to solve are changed slightly. It is also the most direct path toward programming the kind of “common sense” needed not only for problems generally recognized as “hard” but also for the kind that people solve routinely—even unconsciously—in understanding the meanings of ordinary language.
In the past few years, the art of heuristic programming has developed so that writing programs to solve hard technical problems has become almost routine provided that the knowledge required can be put in reasonably restricted forms. Stanford’s DENDRAL and M.I.T.’s MATHLAB programs show this sort of competence. But in all the subjects discussed below, in Learning, Vision, Planning, Natural Language, and Narrative Stories, things are not nearly so clear and the important difficulties all seem to cluster around a few common centers:
a. The problem of organization: hierarchy vs. heterarchy
We find that we can no longer separate programs cleanly into traditional parts and phases like: Pattern Recognition, Induction, Optimization, Tree-Search, Learning, Planning, etc. Instead, as we attack harder and more practical problems, we find that events on one level demand attention at others.
b. The problem of “error”: accidental and deliberate
A second problem, when different kinds and sources of information interact, is that of inconsistency and error. This is not simply a question of reliability—of deciding which is closer to the truth—although that certainly is important. There are deeper reasons why assertions of different kinds cannot always be taken at “face value” at the same level.
c. The problem of collecting and organizing general knowledge
As long as one deals with “toy” problems—puzzles, games, and other situations in which little or no interaction with other aspects of reality are required—Artificial Intelligence techniques are generally quite advanced by human performance standards. But when any understanding of things like space, and time, and people’s desires, and economics, and design, etc., are required, today’s programs hardly approach the competence of a small child.
2. THE PLANNER LANGUAGE
Until recently, most new heuristic programs were written from “scratch”. PLANNER makes available new mechanisms that promise to simplify the now-urgent problems of interlocking all our different kinds of programs.
2.1 Demons
One mechanism PLANNER makes available is instructing it to behave so that whenever X happens, do Y. “X” can be a description in general terms with variables to be assigned, and “Y” can change a goal, add new assertions, erase information, or execute any system procedure. In effect this sets up a “Demon” that lurks in the background until it sees what it is designed to look for.
2.2 Convenience of Many Services
PLANNER offers smooth direct use of the entire LISP system, plus the conveniences of a string-manipulation language, allowing easy handling of subproblem interactions, error escapes, and access to common data bases.
2.3 Automatic Search and Backup Services
In PLANNER, when a subprocess comes to a conclusion it can add this conclusion as an “assertion” to the common data base. Assertions can be stored with “reservations”, so if supporting data is withdrawn later, dependent assertions are automatically withdrawn or relabeled.
2.4 Deductive and Heuristic Assertions
Statements like “A implies B” can be interpreted declaratively or imperatively: “if A is ever asserted, assert B also and add it to the data base”, or as goal advice: “if B is desired as a goal, assert A as a new subgoal to be deduced”.
2.5 Mathematical “Neatness” and Practical Compromise
A simplified implementation, MICRO-PLANNER (Winograd, Sussman, Charniak, Greenblatt), handles the essential pattern-matching and control mechanisms efficiently for current experimental projects.
3. THE VISION PROJECT: TOWARD A HETERARCHICAL PROGRAM ORGANIZATION
Machine Vision serves as an ideal testing ground for studying heterarchical processes. To extend capability from clean geometric blocks to complex real-world scenes, the organizational structure must evolve from a rigid hierarchy to a flexible heterarchy. In heterarchical vision, line-finders, region-growers, shadow-analyzers (ERASER), and body-groupers (SEE) act as communicating modules coordinated by an executive program in PLANNER.
4. NATURAL LANGUAGE AND SEMANTIC INTERPRETATION
Terry A. Winograd has completed a natural language understanding system integrated with robot block manipulation. Built using PROGRAMMAR and MICRO-PLANNER, it carries out syntactic parsing, semantic analysis, deductive reasoning, and physical actions concurrently.
| Sample Dialogue with Winograd’s Natural Language System: Human: PICK UP A BIG RED BLOCK. Computer: OK. Human: GRASP THE PYRAMID. Computer: I DON’T UNDERSTAND WHAT YOU MEAN BY “THE PYRAMID”. (Ambiguous: multiple pyramids exist) Human: FIND A BLOCK WHICH IS TALLER THAN THE ONE YOU ARE HOLDING AND PUT IT INTO THE BOX. Computer: BY “IT” I ASSUME YOU MEAN THE BLOCK WHICH IS TALLER THAN THE ONE I AM HOLDING. OK. Human: WHAT DOES THE BOX CONTAIN? Computer: THE BLUE PYRAMID AND THE BLUE BLOCK. Human: WHY DID YOU CLEAR OFF THAT CUBE? Computer: TO PUT IT ON A LARGE GREEN CUBE. Human: WHY DID YOU DO THAT? Computer: TO STACK UP THE RED CUBE, A LARGE RED BLOCK AND A LARGE GREEN CUBE. Human: WHY DID YOU DO THAT? Computer: BECAUSE YOU ASKED ME TO. |
5. WHY WE ARE STUDYING KNOWLEDGE AND LEARNING
Artificial Intelligence faces a choice between the POWER theory (seeking raw computational speed, resolution theorem-proving, tree search) and the KNOWLEDGE theory (emphasizing epistemological structures, micro-worlds, and common-sense organization). We argue that intelligence stems primarily from rich, well-organized domain knowledge and schematic micro-models rather than pure logical deduction.
6. NARRATIVE, MICRO-WORLDS, AND “UNDERSTANDING”
Understanding stories (such as children’s fables) requires micro-theories of human motives—concepts like FLATTERY, BARGAINING, MISTAKES, LIES, and PURPOSES. Common sense is not a single universal algorithm, but a large, structured catalog of micro-world knowledge, facts, and procedures.
7. MATHEMATICAL THEORY OF COMPUTATION
7.1 Computational Geometry
Developments in continuous perceptron theory (J. Baker) allow rotation and contraction groups to be analyzed with analytical-continuous methods. In line-drawing interpretation, new dual representations and topological line-classification theories (Huffman, Guzman, Hewitt) unify scene analysis mathematics.
7.2 Mathematical Foundations for Theorem-Proving and Problem-Solving
Hewitt and Paterson’s comparative schematology provides formal foundations for PLANNER-like control structures. Complexity limits on resolution theorem-provers highlight the mathematical necessity of integrating domain-specific heuristic knowledge into deductive systems.