The Maryland Large-Scale Integrated Neurocognitive Architecture

Recent progress in neural computation, high performance computing, neuroscience and cognitive science suggests that an effort to produce a general-purpose, adaptive machine intelligence is likely to yield a qualitatively more powerful system than those currently existing. Here we outline our progress in developing a framework for creating such a large-scale machine intelligence, or neurocognitive architecture that is based on the modularity, dynamics and plasticity of the human brain. We successfully implemented three intermediate-scale parts of such a system, and these are described. Based on this experience, we concluded that for the short term, optimal results would be obtained by using a hybrid design including neural, symbolic AI, and artificial life methods. We propose a three-tiered architecture that integrates these different methods, and describe a prototype mini-Roboscout that we implemented and evaluated based on this architecture. We also examined, via computational experiments, the effectiveness of genetic programming as a design tool for recurrent neural networks, and the speed-up obtained for adaptive neural networks when they are executed on a graphical processing unit. We conclude that the implementation of a large-scale neurocognitive architecture is feasible, and outline a roadmap for proceeding.