“…The neuromorphic engineering community has been building physical models of sWTA networks [109], [110], [111], attractor networks [112], [113], and plasticity mechanisms [91] that cover the full range of temporal and spatial scales described in Section IV-A for many years. For example, several circuit solutions have been proposed to implement short-term plasticity dynamics, using different types of devices and following a wide range of design techniques [114], [115], [116], [117], [118], [119]; a large set of spike-based learning circuits have been proposed to model long-term plasticity [120], [121], [122], [123], [77], [124], [125], [126], [127], [128], [91]; multiple solutions have been proposed for implementing homeostatic plasticity mechanisms [129], [130]; impressive demonstrations have been made showing the properties of VLSI attractor networks [112], [113], [23], [4]; while structural plasticity has been implemented both at the single chip level, with morphology learning mechanisms for dendritic trees [131] and at the system level, in multi-chip systems that transmit spikes using the AER protocol, by reprogramming or "evolving" the network connectivity routing tables stored in the digital communication infrastructure memory banks [132], [133]. While some of these principles and circuits have been adopted in the deep network implementations of Section II and in the large-scale neural network implementations of Section III, many of them still remain to be exploited, at the system and application level, for endowing neuromorphic systems with additional powerful computational primitives.…”