I have just started reading about neural networks and I have a basic question. Regarding “initializing” the Hopfield network, I am unable to understand that notion of initialization. That is, do we input some random numbers? or do input a well defined pattern which makes the neurons settle down first time up, assuming all neurons were at state equal to zero, with other stable states being either 1 or -1 after the input.
Consider the neural network below. Which I have taken from HeatonResearch
Glad if someone clears this to me.
When initialising neural networks, including the recurrent Hopfield networks, it is common to initialise with random weights, as that in general will give good learning times over multiple trials and over an ensemble of runs, it will avoid local minima. It is usually not a good idea to start from the same starting weights over multiple runs as you will likely encounter the same local minima. With some configurations, the learning can be sped up by doing an analysis of the role of the node in the functional mapping, but that is often a later step in the analysis after getting something working.