A Parallel Stochastic Number Generator With Bit Permutation Networks

Implementation of Izhikevich neuron based on stochastic computing using a novel inspired Omega‐Flip stochastic number generator

“…In addition, the SC has more advantages in a neural network using stochastic neurons, especially in area and power consumption. For all of the neural networks which use stochastic neurons, just one LFSR is needed 25 . For the multiplication purpose, it is possible to have more multiplier in parallel but with less area overhead, because the described Omega‐Flip's states are able to be shared with other multipliers due to its “choosing specification.” For instance, it is possible to use one multiplexer array (Figure 3 Omega‐Flip network) for 12 neurons, and this sharing decreases area overhead considerably.…”

“…20 The method of generating stochastic number using LFSRs depends on how to use LFSR in SC blocks. Gupta and Kumaresan 26 have introduced the principle of conventional SNGs 25 which includes an n-bit LFSR and an n-bit weighted binary generator. In these SNGs, 2 n -bit stochastic numbers are generated from that weighted binary number.…”

“…For example, an “AND‐gate” is used to implement multiplication of two string‐bit stochastic numbers, which improves area overhead accordingly 20–23 . In addition to AND‐gate, stochastic number generators (SNGs) are required for SC which convert deterministic numbers (binary number) to stochastic numbers 24–27 . Despite simple computing approach and inherit error tolerance, SC has two main shortcomings: long computation time and relatively low accuracy 1,20,24,28–31 …”

“…There are 5 rows for Omega and 5 rows for flip which have to be connected to a specified random number generator. In this Omega‐Flip method, 25 there are some registers that mentioned as k‐bit shifter . The input of k‐bit shifter is connected to LFSR 12 bits , and the output is k‐bits shift if input.…”

“…[20][21][22][23] In addition to AND-gate, stochastic number generators (SNGs) are required for SC which convert deterministic numbers (binary number) to stochastic numbers. [24][25][26][27] Despite simple computing approach and inherit error tolerance, SC has two main shortcomings: long computation time and relatively low accuracy. 1,20,24,[28][29][30][31] In this paper, a digital implementation of Izhikevich neuron using SC approach is proposed.…”