An unsupervised method is presented which permits a set of model neurons, or a microcircuit, to learn low-level vision tasks, such as the extraction of surface depth. Each microcircuit implements a simple, generic strategy which is based on a key assumption: perceptually salient visual invariances, such as surface depth, vary smoothly over time. In the process of learning to extract smoothly varying invariances, each microcircuit maximises a microfunction. This is achieved by means of a learning rule which maximises the long-term variance of the state of a model neuron and simultaneously minimises its short-term variance. The learning rule involves a linear combination of anti-Hebbian and Hebbian weight changes, over short and long time scales, respectively. The method is demonstrated on a hyperacuity task: estimating subpixel stereo disparity from a temporal sequence of random-dot stereograms. After learning, the microcircuit generalises, without additional learning, to previously unseen image sequences. It is proposed that the approach adopted here may be used to define a canonical microfunction, which can be used to learn many perceptually salient invariances.
Domain-invariant similarity activation map metric learning for retrieval-based long-term visual localization
