Creating robust networks

A simple method for introducing robustness in neural networks is proposed. The authors generally assume a feedforward network of three layers of processing units. They have made a study of the activity of hidden units in feedforward back-propagation networks, and can produce a minimal network which can be retrofitted with the desired robustness criteria. This approach of explicitly modifying the network to add robustness has the added advantage of allowing control over the localization within the network of the safety net hidden units. They can then be well separated from the units they are protecting