Complex Learning in Bio-plausible Memristive Networks
The emerging memristor-based neuromorphic engineering promises an efficient computing paradigm. However, the lack of both internal dynamics in the previous feedforward memristive networks and efficient learning algorithms in recurrent networks, fundamentally limits the learning ability of existing systems. In this work, we propose a framework to support complex learning functions by introducing dedicated learning algorithms to a bio-plausible recurrent memristive network with internal dynamics.
... We fabricate iron oxide memristor-based synapses, with well controllable plasticity and a wide dynamic range of excitatory/inhibitory connection weights, to build the network. To adaptively modify the synaptic weights, the comprehensive recursive least-squares (RLS) learning algorithm is introduced. Based on the proposed framework, the learning of various timing patterns and a complex spatiotemporal pattern of human motor is demonstrated. This work paves a new way to explore the brain-inspired complex learning in neuromorphic systems. The emulation of complex spatiotemporal activities of the human brain cortex is central for understanding the learning behaviors, motor processing and cognition 1 . A variety of implementations based on both software and hardware are constantly developed for this purpose. Software-based artificial neural networks (ANNs) on von Neumann computer are widely applied to pattern classification, speech recognition, computer vision, and so on 2,3 . Whereas, software-based simulations occupy large area and consume high power 4 . In this context, various CMOS hardware-based neuromorphic systems have been developed 5-9 . However, the realization of complex learning in such systems is hindered by the lack of suitable device to model the synaptic plasticity 10,11 , until the recent invention of memristor 12-16 . Memristor, characterized by a pinched hysteresis loop of the I-V curve 17 , is a two-terminal nanodevice whose conductance can be gradually modulated under applied voltage or current 18 . This property is reminiscent of the synaptic plasticity 19,20 . In view of its attributes of low power, high speed and easiness to be crossbar integrated 21-25 , memristor has become one of the best candidates for artificial synapse in neuromorphic systems 19,           . Most of the previous memristive works are focused on feedforward architectures 20,36-39 , and a few cases with feedback connections 40-43 . Nevertheless, the feedforward networks lack internal network dynamics 1 and the existing recurrent networks lack efficient learning algorithms 44 , which makes it difficult to support complex dynamics and emulate the learning functions of the brain. Inspired by the biological evidence that the complex learning ability arises from the efficient self-tuning of recurrent connections among neurons 45 , we propose a framework to support complex learning in memristive systems. The requirements of this framework are three-fold: (1) suitable device to model the synaptic plasticity; (2) bio-plausible recurrent neural network with ongoing internal dynamics; (3) efficient learning algorithm to adaptively modulate the synaptic weights. To this end, we emulate the behaviors of excitatory and inhibitory synapses by combining two iron oxide-based memristors with one amplifier, and further build a bio-plausible recurrent memristive