Unsupervised Detection of Illusory Contours

Abstract

We present an unsupervised artificial neural network that is capable of detecting illusory objects in artificial image data. The network has previously been shown [4, 3] to self-organise to identify multiple causes in data distributions where each of the individual causes are independent and may be hidden from direct observation. Learning in the network is completely local and the form of coding that is achieved is sparse. The network is based on a non-linear Principal Components Analysis (PCA) network using a soft positive threshold function as a non-linearity instead of the typical sigmoid (or logistic) function. After the application of the non-linearity in the network we add zero-mean Gaussian noise. The non-linear PCA aspect of this network ensures an coding which is more accessable to human interpretation and the additive noise ensures a sparse representation in which only as many output neurons as are required are used. Illusory contours are detected when the probability of a feature appearing at the input to the network is so high that the most sparse response is to learn an illusory feature between the actual features. By adding lateral connections between the output neurons to provide temporal context in the network, we can achieve a temporal ordering of outputs in response to a particular feature moving across the input space.