Connectionist Models for Learning in Structured Domains

Structured representations are ubiquitous in different fields such as knowledge representation, language modeling, and pattern recognition. Although many of the most successful connectionist models are designed for "flat" (vector-based) or sequential representations, recursive or nested representations should be preferred in several situations. One obvious setting is concept learning when objects in the instance space are graphs or can be conveniently represented as graphs. Terms in first-order logic, blocks in document processing, patterns in structural and syntactic pattern recognition, chemical compounds, proteins in molecular biology, and even world wide web sites, are all entities which are best represented as graphical structures, and they cannot be easily dealt with vector-based architectures. In other cases (e.g., language processing) the process underlying data has a (hidden) recursive nature but only a flat representation is left as an observation. Still, the architecture should be able to deal with recursive representations in order to model correctly the mechanism that generated the observations. The interest in developing connectionist architectures capable of dealing with these rich representations can be traced back to the end of the 80's. Early approaches include Touretzky's BoltzCONS, the Pollack's RAAM model, Hinton's recursive distributed representations. More recent techniques include labeled RAAMs, holographic reduced representations, and recursive neural networks. Today, after more than ten years since the explosion of interest in connectionism, research in architectures and algorithms for learning structured representations still has a lot to explore and no definitive answers have emerged. It seems that the major difficulty with connectionist models is not just representing symbols, but rather devising proper ways of learning when examples are data structures, i.e. labeled graphs that can be used for describing relationships among symbols (or, more in general, combinations of symbols and continuously-valued attributes).

• Algorithms and architectures for classification of data structures.
• Unsupervised learning in structured domains.
• Belief networks for learning structured patterns.
• Compositional distributed representations.
• Recursive autoassociative memories.
• Learning structured rules and structured rule refinement.
• Connectionist learning of syntactic parsing from text corpora.
• Stochastic grammars and their relationships to neural and belief networks.
• Links between connectionism and syntactic and structural pattern recognition.
• Analogical reasoning.
• Applications, including:
• Medical and technical diagnosis: discovery and manipulation of structured dependencies, constraints, explanations.
• Molecular biology and chemistry: prediction of molecular structure folding, classification of chemical structures.
• Automated reasoning: robust matching, manipulation of logical terms, proof plans, search space reduction.
• Software engineering: quality testing, modularization of software.
• Geometrical and spatial reasoning: robotics, structured representation of objects in space, figure animation, layouting of objects.

Marco Gori
DII, University of Siena Via Roma 56, 53100 Siena (ITALY)
Phone: +39 577 263 610
E-mail: marco@ing.unisi.it

Alessandro Sperduti
DI, University of Pisa Corso Italia 40, 56125 Pisa (ITALY)
Phone: +39 50 887 213
E-mail: perso@di.unipi.it