Spaun performing rapid variable creation. The model is shown several input/output examples. It is then given an input and must produce the output. The model has to figure out the pattern in the presented input/output examples in order to solve the task. This requires induction, and more specifically, syntactic generalization. Both of these features have been argued to be central to human cognition. Several authors have argued that no neural models can do this task without implementing a classical architecture (Hadley, 2009; Fodor & Pylyshyn, 1988; Marcus, 2001; Jackendoff, 2002). Spaun does not implement a classical architecture, but can perform the task.
It has been argued by several researchers that neural models of cognition cannot explain a basic feature of human behaviour: the ability to rapidly generalize over syntactically structured input. Task 6, shown here, demonstrates that Spaun is able to perform such generalization. It is shown a series of input/output pairs that bear some relation to one another. Input is stored in one memory (here, 0014), and output is stored in another (here 14). After having seen three examples, Spaun must inductively determine what the relationship is - and it must do this as quickly as people do. To determine if it is successful, Spaun is provided an input it hasn't seen before (here 0074). As shown, it responds correctly, suggesting that it has figured out the underlying structure of the examples.