From robotic toil to symbolic theft: Grounding transfer from entry-level to higher-level categories1

Abstract

A bstract. Neural network models of categorical perception (compression of withincategory similarity and dilation of between-category differences) are applied to the symbol-grounding problem (of how to connect symbols with meanings) by connecting analogue sensorimotor projections to arbitrary symbolic representations via learned category-invariance detectors in a hybrid symbolic/non-symbolic system. O ur nets are trained to categorize and name 50 ´50 pixel images (e.g. circles, ellipses, squares and rectangles) projected on to the receptive eld of a 7 ´7 retina. They rst learn to do prototype matching and then entry-level naming for the four kinds of stimuli, grounding their names directly in the input patterns via hidden-unit representations ('sensorimotor toil'). We show that a higher-level categorization (e.g. 'symmetric' versus 'asymmetric') can be learned in two very different ways: either (1) directly from the input, just as with the entry-level categories (i.e. by toil); or (2) indirectly, fro m Boolean combinations of the grounded category names in the form of propositions describing the higher-order category ('symbolic theft'). We analyse the architectures and input conditions that allow grounding (in the form of compression/ separation in internal similarity space ) to be 'transferred' in this second way from directly grounded entry-level category names to higher-order category names. Such hybrid models have implications for the evolution and learning of language.

Symbol Grounding and the Symbolic Theft Hypothesis

“…They designed a network that had to associate simple pictures with labels. The network architecture was similar to the one used by Cangelosi, Greco and Harnad (2000) and described earlier (Figure 9.1). There were two distinct sensory modalities (retinal and verbal) in the input and output layers, and two hidden layers.…”

“…A recent model by Cangelosi, Greco and Harnad (2000) simulated this overall process of CP, subsequent acquisition of grounded names, and learning of new high-order symbols from grounded ones (grounding transfer). Three-layer feedforward neural networks were used (see Figure 9.1), having two groups of input units: 49 units simulating a retina and 6 units simulating a linguistic input.…”

Evolving spatial and temporal lexicons across different cognitive architectures

“…Steels claimed that the symbol grounding problem was solved (Steels, 2008); however, his solution did not address i) grounding transfer (Cangelosi et al, 2000), ii) dealing with referential uncertainty (Roy, 2002b;Smith et al, 2006), iii) grounding in cognition (Schulz et al, 2011a(Schulz et al, , 2011b) and iv) grounding across different sensors and cognition (Jung and Zelinsky, 2000).…”

“…Symbols can be grounded in other symbols, so long as "terminal" symbols are eventually grounded in sensors. Higher-level symbols can be grounded using combinations of entry-level symbols, in a process called grounding transfer (Cangelosi et al, 2000). Entry-level symbols are first grounded directly in perception, and then higher-level symbols are created by categorizing across entry-level symbols.…”

“…This toponym lexicon can then be used to bootstrap generative conversations, such as how-far and what-direction -conversations that allow the Lingodroids to form a set of lengths and directions corresponding to the distances and angles between toponyms. This bootstrapping involves indirectly grounding the elements, a process known as 'grounding transfer' (Cangelosi et al, 2000).…”