While substantial progress is being made in cognitive robotics, many of us feel that systems now on the horizon will still be a far cry from the original vision of AI as general human-like intelligence acting in the real world. The key limitation may be the lack of autonomy, the ability of a system to behave intelligently in a broad set of situations without the need for specific programming for each task or for a specific learning regime for each new setting. The classical roadblock of linking cognitive architectures to the world through perception, motor control, and background knowledge is being shifted by powerful methods of neurally inspired machine learning and probabilistic reasoning. But, at their core, today's cognitive robots are still driven by conventional algorithmic frameworks that integrate diverse mathematical methods into what resembles classical cognitive architectures. Is that fact a key obstacle to true autonomy of behavior, cognition, and learning?
A radical alternative is to abandon the hybrid approach and seek pervasively neural systems that use no conventional forms of information processing. Such systems would instead generate activation patterns linked to sensors, motors, and neural memory systems and may ultimately be implemented directly in neuromorphic hardware. Early success were achieved for low-level robotic systems that were largely input driven. Recent progress in neural dynamics may unlock cognition for such systems by enabling the generation of activation patterns from within the neural networks, not primarily from input. Dynamic instabilities then generate sequences of activation states to perform cognitive operations.
Our goal in the workshop is to critically discuss roadblocks toward autonomy of neuronal architectures. We will bring together diverse ideas toward a vision of pervasively neural cognitive processing. We will lay out a road-map for such a radical neural vision of autonomous intelligence.