Neuron-centric Hebbian Learning

Figure 1: Graphical representation of the traditional HL (left) and our Neuron-centric Hebbian Learning model (right).Regarding HL, when the network is initialized (first arrow) it starts a loop of observations and actions, in which HL updates the weights (black arrow) by using the equation shown in the red box (see also Equation ( 2)).As can be seen, the parameters of the Hebbian rule in HL are specific to each synapse, leading to a total of 5 parameters to optimize, with being the number of synapses.Alternatively, our NcHL model (right) updates the weights (black arrow) by using the equation shown in the blue box (see also Equation ( 3)).In this case, the parameters of the Hebbian rule are specific to each neuron, hence resulting in a total of 5 parameters to optimize, with being the number of neurons.Assuming ≪ , optimizing NcHL models becomes notably less difficult than optimizing HL models, yet achieving comparable performance.