Figure 2.

(A) One possible implementation of the neuron update pipeline for a conventional Artificial Neural Network (ANN). In the ANN the activation is a multi-bit value (typically FLOAT16 and even INT8 in some recent hardware) which must be multiplied by the connection weight, W_ij before being added to the soma potential (or accumulator value), U_j. (B) In the equivalent SNN implementation the neuron output is a single spike whose arrival causes the associated weight W_ij to be added directly to the potential U_j and no multiplication is necessary.