We are eager to make progress on the general problem of understanding intelligence and building intelligent systems, but we are frustrated because progress in general, and our own in particular, has been slower and narrower than we have expected.}
There has been substantial progress in some of the subfields of Artificial Intelligence during the past three decades, but the field overall is moving toward increasing subfield isolation and increasing attention to near-term applications, retarding progress toward comprehensive theories and deep scientific understanding, and ultimately, retarding progress toward developing the science needed for higher-impact applications.
Accordingly, we have been meeting to see how we might redirect our research toward a program focused on understanding natural intelligence and building integrated intelligent systems. Based on discussions among ourselves and with others, including cognitive scientists and neurobiologists, we have concluded that we can and should redirect our individual research programs, that we should establish a basis for exciting joint work and long-term collaboration, and that we should set down our views so that others can decide if they want to join with us in that joint work and collaboration.
One direction that we find particularly appealing is to study how our vision, language, and motor faculties contribute to overall intelligence and how those faculties enable humans to capture and exploit enormous quantities of knowledge, both commonsense and domain specific. In consequence of that understanding, we will enable ourselves and others to build much more powerful language systems, vision systems, and knowledge bases.
We think it is informative to ask how similarities in brain organization provide a substrate for analogous computational activity. Each modality draws on proprietary representations and problem-solving competences. Each takes inputs from internal sources as well as from the outside world. A linguistically represented problem can require us to harness the power of our visual system so as to engage the strategies of "visual imagination." An important contribution to high-level cognitive ability emerges from cooperative problem-solving as peripherals work out parts of a problem that have been reexpressed in terms of peripheral-specific representations. This cooperative work engages sophisticated mechanisms for actuating, correlating, synchronizing, caching, remembering, and re-experiencing.
Neuroscience teaches that the various modalities are implemented with similar "wetware." Different regions of the neocortex subserving different functions and modalities have similar laminar organization and connectivity patterns. Each modality consists of multi-layers interconnected so that if any one area gets inputs from another, it sends back outputs to that area. The systems supporting the various modalities are multiply and reciprocally cross linked. The organization provides a natural substrate for a close, bidirectional interaction between more peripheral and more central processes. Such bidirectional interaction can simultaneously transform stored prototype models and raw sensory signals for the purpose of integration, matching, and recognition.
Traditional approaches to Artificial Intelligence have been largely organized around unconstrained algorithms or simplistic biological models. In the past, both lack of constraints and incorrect constraints have imposed show-stopping penalties. Today, by contrast, the perspective of computer science, tempered by hardware insights accumulating in neurobiology, can provide new direction to the enterprise of understanding intelligence. Stimulation should, of course, flow in both directions.
In summary, we believe the time is ripe to refocus our thinking and work toward a ever-more-correct series of computational ideas, unified architectures, and experimental programs. We invite others who have reached similar conclusions to join with us.