By Peter Voss | 15 April 2017
What we need is a cognitive architecture (or autonomous agent) approach. However, even this by itself is not enough: essential cognitive mechanisms (such as knowledge & skill representation, short-term memory & context, reasoning & planning, perception & action, Focus & selection, metacognition, etc.) need to be tightly integrated. The standard engineering approach of separate (and disparate) modules cannot achieve this objective.
Pat Langley has an excellent essay in which he highlights differences between the current mainstream approaches to AI, and what he calls the ‘Cognitive Systems Paradigm’:
In this essay, I review the motivations behind the cognitive systems movement and attempt to characterize the paradigm. I propose six features that distinguish research in this framework from other approaches to artificial intelligence, after which I present some positive and negative examples in an effort to clarify the field’s boundaries.
Here’s my summary interpretation of these points:
1. High-Level Cognition: Abilities such as abstract reasoning, deep comprehension, goal-directed planning, and problem solving. Contrast this with (statistical) pattern recognition, classification, and prediction—what is known as machine learning.
2. Structured Representations: All knowledge and skills (from perception/ action all the way to symbols) should be encoded in a uniform, integrated manner way that reflects the logic structure of the data, and is somewhat scrutable. Mainstream AI approaches such as machine learning and databases tend to not be uniform, integrated, or to match the structure and flexibility of the data.
3. System-Level Approach: Comprehensive, integrated cognitive architectures provide seamless integration between various cognitive functions—these functions are thus mutually supportive. Current AI approaches tend to take the opposite approach, combining separate, specialized modules via pipelines or narrow APIs.
4. Heuristics and Satisficing: Systems should be designed to be practical with incomplete (or even incorrect) information, and with limited time or computing resources. Generally, learning and reasoning systems that are not designed as interactive, real time, cognitive architectures are unable to cope with ambiguity and severe time constraints (e.g. immediate learning).
5. Links to Human Cognition: While AI designs do not have to copy how brains achieve intelligence, crucial characteristics of human cognition must be implemented in a successful, practical intelligence engine. These include the ability to handle ambiguity, abstract conceptualization and reasoning, short-term memory and context, as well as meta-cognition. Typical machine learning, conversational agents, and robotics does not generally implement these features.
6. Exploratory Research: Ideally, research towards general-purpose, human-level AI should be ruthlessly focused on general aspects of intelligence, and not specific, narrow applications or algorithms. At the same time, R&D should be guided by practical experimental results and not overly by academic or theoretical consideration—i.e. more engineering than theory.
Reprinted with permission from the author.
— Church and State (@ChurchAndStateN) April 8, 2018
Peter Voss, AGI Innovations inc @ iHuman: The Future of Minds and Machines, SVForum
Peter Voss on Applying Artificial Intelligence to Intelligence Enhancement and Life Extension
Erica: Man Made
Be sure to ‘like’ us on Facebook