Most unsupervised neural networks perform some form of gradient descent on an error
function, normally the sum squared error, and as such are likely to fall into a local
minimum. Attempts to overcome this problem have concentrated on two approaches:
modifying the objective function and thereby the update rule, and secondly using some
form of simulated annealing [Metropolis et al., 1953]. In practice, in both approaches,
all the nodes in the network are updated on each vector presentation, rather than just the
winner. The term soft competition is used to embrace both methods.