Cognition breaks through the 'here-and-now barrier', using memory to anticipate the future, learn and develop. By virtue of this development, cognition also breaks through the 'a priori knowledge' building the system's own understanding of its environment. Of course, the obvious question that arises is: How does it do this?

My personal position is that a cognitive system is an autonomous anti-entropy engine, a system that is capable of creating order locally. Specifically, cognition is the dynamic evolution of a complex non-linear re-entrant self-organizing system that decreases its entropy as it survives environmental perturbations & maintains its autonomy. Thus, there are two complementary processes going on in a cognitive system. One is the self-organization, which maintains the autonomy of the system. This is probably achieved in some sort of process of autopoiesis (or self-production, after Maturana and Varela). The other is a process of self-development, whereby the system changes its organization so as to increase its space of possible interaction and, consequently, decrease the entropy of some function of its organization and behaviour in the face of environmental perturbations (which may well include interactions with other cognitive agents). These complementary processes are captured nicely by the two circles in Maturana's and Varela's ideogram of two interacting cognitive systems show below.

Autonomy implies a self-governing self-organizing system. Hence, it is not controlled by external influences, although it may of course react and adapt to external influences. That is, external influences are perturbations, not control signals. Again, this idea is due to Maturana and Varela.

Anti-entropy implies that, over time, the system dynamics change so that there is an increase in the order of some characteristic of the system's dynamics. And, again, this idea corresponds to Maturana's and Varela's distinguishing property of a cognitive system: the ability to self-modify. So what is this characteristic? I postulate that it is the interactions of the system with its environment, including other agents also including itself. In other words, the space of possible interactions increases over time and as entropy of interaction decreases as a function of its history of interaction, i.e., the system becomes more robust to perturbations. There is no requirement that this entropy reduction (qua increase in the space of interaction) occur in a monotonic fashion. It often happens non-monotonically, e.g. such as is the case when children learn to walk and when adults learn a new skill or language; you get worse before you get better.

This entropy reduction manifests itself as (a) an ability to adapt to new perturbations or stimuli, and (b) an ability to anticipate or predict the consequences of the system's actions. It is this anticipatory capacity that facilitates the non-monotonic development: it allows the system to effect some prospective projection of what might be in order to guide it through the developmental period.

So, as noted above, cognitive systems adapt and anticipate, but they adapt to enhance and improve the anticipatory capability, and the anticipatory ability, in turn, facilitates the adaptation.

That's my position on cognition. What about perception, then? My view of perception dovetails with cognition: Perception is a system-dependent process that acts to filter and shape the morass of stimulus perturbing the system with the express purpose of facilitating the effective action of that system (i.e. maintain its autonomy through self-organization and decrease its entropy though self-development). Thus perception is dependent on the cognitive system; it is not unique mapping from a fixed external world to internal representation but instead a process of system dependent structuring. Berkeley was one of the first to articulate this view of perception in his Essay Towards a New Theory of Vision but, ultimately, this is a phenonemological view of perception and cognition (in the strict sense of Husserl's and Heidigger's philosophy). This structuring can be viewed as a sort of transform (or set of transforms), where the transform is specified by the system's own particular state of organization and development. What is peculiar to all transform techniques - Fourier, Mellin, Laplace, Hough, whatever - is that they map from one space where information is distributed and implicit to one where relevant information is localized and explicit, and therefore accessible and useful (in the sense of facilitating the system's autonomy and self-organization). How are these transforms identified? Some are certainly present in the phylogenetic configuration of the system but others arise from the ontogenetic process, that is, as a consequence of the self-development of the system. In other words, a system's perceptual processes arise from its cognition and these perceptual processes facilitate cognition. If this all sounds very circular, that's exactly what you would expect because, as many researchers such as Edelman, Bickhard, and of course Varela, have indicated over the years, cognitive processes are strongly entwined with the concepts of recursion and reentrance.

My research programme for the next X years is to validate this position empirically by writing software for the iCub, beginning with its cognitive architecture.