From robotic toil to symbolic theft: grounding transfer from entry-level to higher-level categories

Neural network models of categorical perception (compression of within-cate gory similarity and dilation of between-category differences) are applied t o the symbol-grounding problem (of how to connect symbols with meanings) by connecting analogue sensorimotor projections to arbitrary symbolic represe ntations via learned category-invariance detectors in a hybrid symbolic/non -symbolic system. Our nets are trained to categorize and name 50 x 50 pixel images (e.g. circles, ellipses, squares and rectangles) projected on to th e receptive field of a 7 x 7 retina. They first learn to do prototype match ing and then entry-level naming for the four kinds of stimuli, grounding th eir names directly in the input patterns via hidden-unit representations (' sensorimotor toil'). We show that a higher-level categorization (e.g. 'symm etric' versus 'asymmetric') can be learned in two very different ways: eith er (1) directly from the input, just as with the entry-level categories (i. e. by toil); or (2) indirectly, from Boolean combinations of the grounded c ategory names in the form of propositions describing the higher-order categ ory ('symbolic theft'). We analyse the architectures and input conditions t hat allow grounding (in the form of compression/separation in internal simi larity space) to be 'transferred' in this second way from directly grounded entry-level category names to higher-order category names. Such hybrid mod els have implications for the evolution and learning of language.