How Can a Massively Modular Mind Be Context-Sensitive? A Computational Approach Giovanni Pezzulo (giovanni.pezzulo@istc.cnr.it) Istituto di Scienze e Tecnologie della Cognizione - CNR Via San Martino della Battaglia, 44 - 00185 Roma, Italy Abstract Starting from a question by Dan Sperber, we analyze the expressive power of massively modular systems from a computational point of view. The key operation is introducing a non-computational component, a limited amount of resources to be shared among the modules. A case study is provided showing that such systems are flexible and context-sensitive, also having interesting cognitive features such as priming and memory effects. Introduction Massive modularity is the hypothesis that in the mind there are not central processes, but all its cognitive functionalities are realized by modules of all format and size, implementing domain-specific abilities. Fodor (2000) challenges this view: in his view only periferic processes are modular; higher cognitive processes are context sensitive and flexible, two features that can not be obtained by using encapsulated modules performing local computations, but only by central, non-modular processes. Sperber (2005) replies that Fodor consid- ers only the computational operations performed inside modules, that are local and not context sensitive; but there is another possible solution: to add a non compu- tational element, resources, and treat their dynamics as “influences” between modules. Adopt a strong modular- ist view of the mind, [...] but assume also that each such process takes resources, and that there are not enough resources for all processes to take place. All these poten- tially active modules are like competitors for resources. It is easy to see that different allocations of resources will have different consequences for the cognitive and epis- temic efficiency of the system as a whole (Sperber, 2005). In the next sections we review the main requirements for adding flexibility to a massively modular system. We also introduce AKIRA, a multi-agent framework where agents, implementing modules, share and exchange a limited amount of resources. Last, we design a simple massively modular system in AKIRA and we test it on simple tasks requiring context sensitiveness. Modularity and Massive Modularity Modularity has been widely accepted, even if there is debate about how it is realized in biological systems. Fodor (2000) argues in favor of “horizontal modular- ity”, i.e. implementing whole cognitive processes such as perception, attention and memory as modules. On the contrary, massive modularity claims that cognitive processes emerge from the interactions between many semi-independent, concurrent, domain-specific compo- nents. In AI, the Society of Mind model (Minsky, 1986) analyzes cooperating and competing processes realizing complex cognitive operations: we call this “vertical mod- ularity”. Horizontal modules are often impenetrable: this hap- pens for example for the perceptual-motor and mem- ory modules in cognitive architectures such as ACT-R. While inside each module there is interaction and com- petition between representations and processes, there is limited interaction between modules. Blendings, multi- modal interactions and other contextual and interference effects are only possible at a coarse-grained level between modules’ outputs (that have also to share the same for- mat, e.g. symbolic). For instance, it is possible that the output of a module facilitates or inhibits another module, but this can not occur at a more fine-grained scale: representations and processes in the modules are independent and no module assumes as contextual pa- rameters the current activity of the other ones. We pro- pose instead that this should occur, since many cognitive processes are likely to have place thanks to -and in the context of- each other, at least to a certain extent. More- over, in horizontal modules all influences are mediated by representations; we propose instead an additional, non computational element, resources, permitting to repre- sent indirect pressures. Context Sensitivity and Flexibility Context-sensitivity and flexibility (intended as context- sensitivity on a longer time scale) can be introduced in modular systems by following some design principles; we propose two about representations: modules should al- low introspection and synchronize some of their inner states and two about non computational elements: mod- ules should compete for limited resources and be able to influence the priority of other modules. Introspection and Synchronization. There are many cases where modules have to make available some of their content to other ones. This feature can be ex- ploited for meta-reasoning, where some processes are supposed to have access to the content and the processes of other modules. But it is also very useful for building hierarchies, where some modules exploit the results of other ones. In a sense this is the opposite of encapsula-