Cite as:
Hyvarinen A, Hoyer P, 1999, "Emergence of complex cell properties in a neural network that maximises the sparseness of local energies" Perception 28 ECVP Abstract Supplement

Emergence of complex cell properties in a neural network that maximises the sparseness of local energies

A Hyvarinen, P Hoyer

A useful approach to modelling the properties of visual neurons is based on statistical generative models of natural images. Olshausen and Field [1996 Nature (London) 381 607 - 609)] showed that the estimation of a simple linear generative model can be performed by maximising the sparseness of the underlying image components. Sparseness is a statistical property that expresses the 'spiky' non-Gaussian shape of the probability density function of the component. Olshausen and Field used natural image patches as training data (input) to a neural network that learned according to this estimation principle, and observed emergence of linear features that closely resemble simple-cell receptive fields. We show here that this same principle can explain the emergence of the principal properties of complex cells as well. We modelled complex-cell responses using classical (local) energy models, and derived a learning rule for a neural network to maximise the sparseness of the responses. We trained the neural network using 16 × 16 pixel monochrome image patches from natural scenes as the input data. Thus we obtained features that had the principal properties of complex cells: phase and (limited) shift invariance, in addition to orientation and frequency selectivity.

These web-based abstracts are provided for ease of seaching and access, but certain aspects (such as as mathematics) may not appear in their optimum form. For the final published version of this abstract, please see
ECVP 1999 Abstract Supplement (complete) size: 1478 Kb