Dynamic liquid association: Complex learning without implausible guidance

Simple associative networks have many desirable properties, but are fundamentally limited by their inability to accurately capture complex relationships. This paper presents a solution significantly extending the abilities of associative networks by using an untrained dynamic reservoir as an input filter. The untrained reservoir provides complex dynamic transformations, and temporal integration, and can be viewed as a complex non-linear feature detector from which the associative network can learn. Typically reservoir systems utilize trained single layer perceptrons to produce desired output responses. However given that both single layer perceptions and simple associative learning have the same computational limitations, i.e. linear separation, they should perform similarly in terms of pattern recognition ability. Further to this the extensive psychological properties of simple associative networks and the lack of explicit supervision required for associative learning motivates this extension overcoming previous limitations. Finally, we demonstrate the resulting model in a robotic embodiment, learning sensorimotor contingencies, and matching a variety of psychological data.