Smolensky is the recipient of the 2005 Rumelhart Prize
for his development of the ICS Architecture, a model of cognition that aims to unify connectionism
, where the symbolic representations and operations are manifested as abstractions on the underlying connectionist or artificial neural networks
. This architecture rests on Tensor Product
compositional embeddings of symbolic structures in vector spaces. It encompasses the Harmonic Grammar
framework, a connectionist-based numerical grammar formalism he developed with Géraldine Legendre
and Yoshiro Miyata,
which was the predecessor of Optimality Theory
. The ICS Architecture builds on Harmony Theory, a formalism for artificial neural networks
that introduced the restricted Boltzmann machine
architecture. This work, up through the early 2000s, is presented in the two-volume book written with Géraldine Legendre, The Harmonic Mind
Subsequent work introduced Gradient Symbolic Computation
, in which blends of partially-activated symbols occupy blends of positions in discrete structures such as trees or graphs.
This has been successfully applied to numerous problems in theoretical linguistics where traditional discrete linguistic structures have proved inadequate,
as well as incremental sentence processing in psycholinguistics.
In work with colleagues at Microsoft Research and Johns Hopkins, Gradient Symbolic Computation has been embedded in neural networks using deep learning to address a range of problems in reasoning and natural language processing.
Among his other important contributions is the notion of local conjunction of linguistic constraints, in which two constraints combine into a single stronger constraint that is violated only when both of its conjuncts are violated within the same specified local domain. Local conjunction has been applied to the analysis of various "super-additive" effects in Optimality Theory. With Bruce Tesar (Rutgers University
), Smolensky has also contributed significantly to the study of the learnability of Optimality Theoretic grammars (in the sense of computational learning theory