Fig. 4: GRNG-based synapse and modified tanh activation function.

a Schematic of the GRNG-based synapse. The input to the synapse, \({V}_{{{{{{\rm{in}}}}}}}\) is applied as +\({V}_{{{{{{\rm{in}}}}}}}\) and -\({V}_{{{{{{\rm{in}}}}}}}\) to the memtransistors, \({{{{{{\rm{T}}}}}}}_{+}\) and \({{{{{{\rm{T}}}}}}}_{-}\) with conductance \({G}_{+}\) and \({G}_{-}\) (modulated using \({V}_{{G}_{+}}\)and \({V}_{{G}_{-}}\)), respectively. The effective conductance of this synapse is given by \({G}_{{{{{{\rm{eff}}}}}}}={G}_{+}-{G}_{-}\), allowing positive and negative conductance. b Waveform applied to the synapse to generate GRNs with independent control over \(\mu\) and \(\sigma\). c \({G}_{+}\), -\({G}_{-}\) and \({G}_{{{{{{\rm{eff}}}}}}}\) of the synapse sampled 300 times, where \({{{{{{\rm{T}}}}}}}_{+}\) controls \(\sigma\) and \({{{{{{\rm{T}}}}}}}_{-}\) controls \(\mu .\) GRNG-based synapse showing d independent control of \({\mu }_{{G}_{{{{{{\rm{eff}}}}}}}}\) for constant \({\sigma }_{{G}_{{{{{{\rm{eff}}}}}}}}\) and e independent control of \({\sigma }_{{G}_{{{{{{\rm{eff}}}}}}}}\) for constant \({\mu }_{{G}_{{{{{{\rm{eff}}}}}}}}\). f Linear scaling of the synaptic output (\({I}_{{{{{{\rm{out}}}}}}}\)) distribution as the function of \({V}_{{{{{{\rm{in}}}}}}}\). g Schematic of circuit for the modified tanh activation function using a n-type MoS₂ memtransistor (\({{{{{\rm{T}}}}}}1\)) and a V-doped p-type WSe₂ memtransistor (\({{{{{\rm{T}}}}}}2\)), where the input voltage (\({V}_{{{{{{\rm{S}}}}}}}\)) is applied to the gate terminal of \({{{{{\rm{T}}}}}}1\) and \({{{{{\rm{T}}}}}}2\). h The transfer characteristics of the circuit (solid line) i.e., output voltage (\({V}_{{{{{{\rm{O}}}}}}}\)) versus \({V}_{{{{{{\rm{S}}}}}}}\), closely models the tanh activation function (dotted line).