A software I'm using has implemented two unsupervised learning algorithms, Oja's and Bienenstock, Cooper, Munro's (BCM) learning rule. I understand that they are two very different algorithms for adjusting a neuron's connection weight, but I was wondering what is the relation between them. Do they both have the same plausibility? Can they be used together or are they mutually exclusive?
Although the effects of BCM and Oja's learning rule have been explored superficially in a computational manner in the context of the NEF, judging from the results (which I've reproduced below) it isn't clear to me if they can/should be combined in any meaningful way.
The best hypothesis that I can come up with is BCM forces the ensemble to be either "ON" or "OFF" by minimising transitory states, while Oja provides limits to how much the connection weights can change. So in a way, Oja provides bounds to BCM? Further investigation is needed.