| Abstract | The idea of creating a general purpose machine intelligence that capturesmany of the features of human cognition goes back at least to the earliest days
of artificial intelligence and neural computation. In spite of more than a
half-century of research on this issue, there is currently no existing approach
to machine intelligence that comes close to providing a powerful, general-purpose
human-level intelligence. However, substantial progress made during recent years
in neural computation, high performance computing, neuroscience and cognitive
science suggests that a renewed effort to produce a general purpose and adaptive
machine intelligence is timely, likely to yield qualitatively more powerful
approaches to machine intelligence than those currently existing, and certain
to lead to substantial progress in cognitive science, AI and neural computation.
In this report, we outline a conceptual framework for the long-term development
of a large-scale machine intelligence that is based on the modular organization,
dynamics and plasticity of the human brain. Some basic design principles are
presented along with a review of some of the relevant existing knowledge about
the neurobiological basis of cognition. Three intermediate-scale prototypes for
parts of a larger system are successfully implemented, providing support for the
effectiveness of several of the principles in our framework. We conclude that a
human-competitive neuromorphic system for machine intelligence is a viable long-
term goal, but that for the short term, substantial integration with more
standard symbolic methods as well as substantial research will be needed to make
this goal achievable.
|
