Emergent componential coding of a handwritten image database by neural self-organisation

This paper demonstrates the unsupervised discovery of localised components in real image data, using images of much larger size than the small fragments from which components have previously been extracted. The handwriting images used are also much more homogeneous than the random natural scenes used in earlier demonstrations, containing components of a specific size-scale and structure. Because of this homogeneity, the components found are not wavelets covering a range of size scales: instead, they correspond to line- and curve-segments made by the pen. The objective function that is optimised here encodes and reconstructs the data via a Markov process, and is also related to density modelling techniques. Several earlier theoretical and experimental results can also be attributed to the form of neuron used here, including the extraction of words from continuous speech and the discovery of unknown transformation invariances via the controlled breaking of dynamical symmetry.