The AI community has cause for much pride in the progress it has made over the past half century. We have made significant headway in solving fundamental problems in representing knowledge, in reasoning, in machine learning, and more. On the practical side, AI methods now form a key component in a wide variety of real-world applications. However, a solution to the AI problem—achieving a truly intelligent system—remains elusive.
The capabilities that appear the hardest to achieve are those that require interaction with an unconstrained environment: machine perception, natural language understanding, or common-sense reasoning. To build systems that address these tasks, we need to draw upon the expertise developed in many subfields of AI. Of course, we need expertise in perception and in natural language models. But we also need expressive representations that encode information about different types of objects, their properties, and the relationships between them. We need algorithms that can robustly and effectively answer questions about the world using this representation, given only partial information. Finally, as these systems will need to know an essentially unbounded number of things about the world, our framework must allow new knowledge to be acquired by learning from data. Note that this is not just “machine learning” in its most traditional sense, but a broad spectrum of capabilities that allow the system to learn continuously and adaptively.
Therefore, in addition to making progress in individual subfields of AI, we must also keep in mind the broader goal of building frameworks that integrate representation, reasoning, and learning into a unified whole.