Soft bodies as input reservoir: role of softness from the viewpoint of reservoir computing

Biological systems are composed of a large number of continuum elements and they effectively realize their flexible and smooth functionalities by capitalizing on the rich dynamics occurring in their soft bodies. In particular, the diverse spatiotemporal patterns of the continuum body have recently been demonstrated as highly useful in implementing a certain class of computation, implying that soft bodies can work as computational devices like nervous systems. However, unlike nervous systems whose activities are basically chaotic and spontaneous, soft bodies are passive and eventually stop their movements without undergoing external force. In addition, both nervous systems and soft bodies are tightly coupled in biological systems and seamlessly contribute to establishing the functionalities as a whole, making it difficult to clarify the specific roles of each element and the degrees of contribution. In this paper, we quantitatively investigate a general role of the soft bodies by introducing a new concept called
_{input
reservoir}
and examining the contribution of the passive elements to information processing. Input reservoir refers to a dynamical system that acts as a filter with a fading memory property whose output is projected onto a chaotic dynamical system and works as an interface between the external stimulus and the chaotic system. Soft bodies can be considered as one type of input reservoir in terms of function because soft bodies are placed between the environment and the chaotic nervous system. We show that the performance of a certain type of temporal task can be improved by installing an input reservoir and training the internal weights of the connected chaotic neural network using a reservoir computing method called
_{innate
training}
. Moreover, we demonstrate that complicated spatiotemporal patterns collected by sensing the soft body dynamics significantly contribute to increasing the computational capacity of the whole by projecting them onto a chaotic neural network and modifying the network weights by innate training. These results imply that soft continuum elements in biological systems have a significant effect in the system's information processing even when the soft elements are firmly joined with the nervous system.