AIM Networks : Autolncursive Memory Networks for Anticipation Toward Learned Goals

The ability to anticipate future states is a key adaptive property of living systems (Glenberg, 1997). Robert Rosen (1985) suggested that an anticipatory system is characterized by finality, and "is a system containing a predictive model of itself and/or of its environment, which allows it to change state at an instant in accord with the model's predictions pertaining to a later instant". Daniel Dubois (Dubois & Resconi, 1992; Dubois, 1998a, 2000) defined the concept of incursive and hyperincursive anticipatory systems, able to generate respectively one or several anticipations influencing the computing of the next state of the system. In this article, the concept of autoincursion is proposed as the ability for a system to compute its successive internal states as a function of its past, present and anticipated states, to select among several anticipated states, and to autonomously change its own equation parameters by learning. Some fundamental properties of a neural network architecture and dynamics are proposed to define Autolncursive Memory Networks. AIM Networks can learn and activate multiple attractors simultaneously, exhibiting synergic dynamics of attractors encoding external inputs. This allows them (l) to compute their successive states as a function of past, present, and multiple anticipated states, (2) to change the way they compute their successive states through symmetric or asymmetric modification of the synaptic structure during autonomous leaming, and (3) to select sequences of anticipations oriented toward learned goals.