Contents lists available at ScienceDirect Solid State Electronics journal homepage: www.elsevier.com/locate/sse Electrical Tunability of Partially Depleted Silicon on Insulator (PD-SOI) Neuron Sangya Dutta a, ⁎ , Tanmay Chavan a , Nihar R. Mohapatra b , Udayan Ganguly a, ⁎ a Department of Electrical Engineering, Indian Institute of Technology Bombay, Powai, Mumbai 400076, India b Department of Electrical Engineering, Indian Institute of Technology Gandhinagar, Gujarat 382355, India ARTICLEINFO Keywords: PD-SOI LIF neuron Impact ionization Electrical tunability Threshold Sensitivity SNN ABSTRACT The hardware realization of spiking neural network (SNN) requires a compact and energy efcient electronic analogtothebiologicalneuron.Aknobtotunetheresponseoftheas-fabricatedneuronallowsthenetworkto performvariousfunctioningwithoutalteringthehardware.Earlier,ourgrouphasexperimentallydemonstrated anLIF(leakyintegrate&fre)neurononahighlymatured32nmSOICMOStechnology.Inthiswork,wehave experimentally demonstrated electrical tunability of the same through its intrinsic charge dynamics based on impactionization(II)enabledfoatingbodyefect.First,atunableinputthreshold( V th )isachievedbychanging the drain bias. Second, above threshold, a fring frequency (f) to input (V) sensitivity (df/dV) tuning is suc- cessfullydemonstratedbycontrollingtheSOI-MOSFET’scurrentthreshold.Weshowthatboththeindependent controlofsensitivityandthresholdisfundamentallyenabledbythenon-linearityoftheimpactionizationbased carrierdynamics.TheSOIneuronprovidesequivalentelectricaltunabilitytoResistor-Capacitor(RC)basedLIF neurons without degrading its original area and power advantages for clock-less, asynchronous SNNs. Further, we show that the neuronal behavior (threshold and sensitivity) is a key determinant of network performance, specifcallythelearningaccuracy.Such fexibilitybasedonpost-fabricationelectricaltuningwillbeanattractive enabler for the SNN hardware. 1. Introduction Spiking neural network (SNN) is a biologically plausible neural network [1] which requires thorough design for its efcient func- tioning.Typically,thedesignschemeconsistsofthechoiceofneuron/ synapse model, choosing the right learning algorithm, selection of a compact and energy efcient network architecture [2]. However, a crucialchallengeliesinthehardwarerealizationofsuchasystem [3]. Neuronal design is a specifc challenge that requires that tuning of neuronal behavior or adding biological features [4]. A neuron’s beha- viorisgenerallymanifestedinitsresponsecurve(outputfrequencyvs. input stimulus). As shown in Fig. 1a, two important parameters that evaluate neuronal response are the (i) input threshold and (ii) the sensitivity (slope of the response curve i.e.df dV / ). The threshold vol- tage( V th )isdefnedbytheminimuminputstimulus( V )beyondwhich the neuron shows a fnite frequency ( f ) i.e. > f 0 when V V th . The sensitivity is the rate of change in the frequency w.r.t the input above threshold i.e. ( df dV / at > V V th ). Thus, tuning a neuron’s character- istics primarily requires modifcation of these two parameters. A tun- ablethresholdenablesselectiveresponsetoacertainrangeofinput.For example, inputs below threshold ( < V V th ) are rejected, a feature that may help in fltering the input [5]. A sharp nonlinear frequency re- sponseresultsinbettersensitivity.Incertainapplications,e.g.,contour tracking during chemotaxis [6], excessive sensitivity may produce in- accurate tracking due to oscillation about the target contour. Thus, sensitivity tuning is attractive. Therefore, if electrically tunable, these parameterscancompletelymodifythebehaviorofneuronsenablingthe neural network to execute various applications. Although several groups have demonstrated electronic neuron on Si [7–10,4], platform, most of these hardware neurons do not demonstrate electrical tun- ability apart from the Diferential-Pair Integrator (DPI) neuron [4]. However, the DPI neuron adds to the area/power consumption to en- able such tunability unlike SOI neuron. Among the non Si-based arti- fcial neurons [11–14], PCMO based IF (integrate & fre) [14] neuron shows a similar capability by modulating the input pulse width. How- ever,theseneuronsrequirepulsedinputs.Suchaconstraintisunableto support continuous input, which is required for more biologically rea- listic asynchronous, clock-less analog SNN [15–17].Tothebestofour knowledge,noothernon-Sineuronshavedemonstratedsuchtunability. Earlier, our group has experimentally demonstrated LIF (Leaky https://doi.org/10.1016/j.sse.2019.107623 Received 20 January 2019; Received in revised form 2 June 2019; Accepted 15 June 2019 ⁎ Corresponding authors. E-mail addresses: sangyadutta@gmail.com (S. Dutta), udayan@ee.iitb.ac.in (U. Ganguly). Solid State Electronics 160 (2019) 107623 Available online 17 June 2019 0038-1101/ © 2019 Elsevier Ltd. All rights reserved. T