Introducing invariance: a principled approach to weight sharing

The paper describes a framework for addressing the training problem of multi-layer perceptrons by a principled introduction of weight sharing. The technique not only reduces the size of the class from which the learning algorithm must select its hypothesis but also reduces the number of examples required for a given level of generalization. The question of assessing the functionality of the weight sharing network is addressed, with a view to ensuring that the weight constraints introduced have not excluded the target functions of the learning task.< >