Despite therapidproliferationofdeeplearninginvariouslarge-scaleapplicationsoverthepast decade, deploymentattheedgeremainslimited.Whileedgecomputingo!ers numerousadvantages, including privacypreservationandlatencyreduction,italsoposessignificantchallengesdue to resourceconstraints,particularlyintermsofenergyandcompute.Theprevailinglearning infrastructureforartificialneuralnetworksonvonNeumann-basedhardwareisnotoptimizedfor suchscenarios.Incontrast,neuromorphiccomputingo!ers analternativeparadigmbyemulating the spikingbehaviorofbiologicalneuronsinsilico,utilizinganalogsignalsforcomputationand digital binaryeventsforcommunication.Theresultingmixed-signaldevicesnativelyimplement Spiking NeuralNetworks(SNN),withthepotentialforon-chiplearningandin-memorycomputing, therebysignificantlyreducingpowerconsumption. Despite thefactthatneuromorphiccomputingwasfirstintroducedmorethanthreedecadesago, training multi-layernetworkson-deviceremainschallenging,ultimatelylimitingtheusabilityfor real-worldtasks.AprimaryreasonforthisisthatexistinglearningframeworksforSNNsareeither limited toshallowarchitecturesorrequiregradientestimationonconventionalhardware,neglecting the spike-basednatureofneuromorphicprocessors. This workaddressesthisgapbyimplementinganovelspike-basedsupervisedlearningalgorithm on aneuromorphicprocessor,theDYNAP-SE.Inthisframework,pairsofcontrolneuronsare added foreveryneuronintheoutputlayerofafeedforwardSNN.Thesecontrolneuronsreceive the networkoutputandatargetandusethemtocomputealocalerrorsignal.Throughrecurrent connections, thecontrollersteersnetworkactivitytowardadesiredtargetwhileupdatingsynaptic weightsusingonlyspike-basedinformation,thereforefulfillingthecriteriaforonline,on-device learning. The algorithmisimplementedontheDYNAP-SEusinganin-the-loopsetupandcarefullycalibrated to accountforconstraintssuchasnoise,substratemismatch,limitednumberofneurons,and quantized,boundedsynapticweights.Toevaluatetheapproach,SNNsaretrainedondi!erent classification tasks.Forsingle-layernetworks,thehardwareimplementationcanexactlymatchthe performanceofsimulatedspikingneuronsonconventionalhardware.Formulti-layernetworks,we observeasimilarqualitativeimprovementandonlyasmallquantitativeperformancedecrease.This workunderscoresthecapacityofthealgorithmtoaddressthespecificconstraintsofmixed-signal devices,leveragingtheavailableinformationforon-chiplearning.Itthereforepavesthewayforthe nextgenerationofenergy-e"cient neuromorphiccomputingandonlinelearningattheedge.