Electronic neuromorphic system, synaptic circuit with resistive switching memory and method of performing spike-timing dependent plasticity
Data di pubblicazione
Data di priorità
Domanda di brevetto negli Stati Uniti
Politecnico di Milano
DIPARTIMENTO DI ELETTRONICA, INFORMAZIONE E BIOINGEGNERIA
Daniele Ielmini, Simone Balatti, Stefano Ambrogio, ZhongQiang Wang
The electronic neuromorphic networks are implemented to reproduce brain-like processing applications wherein principles of computation based on pattern learning and recognition are performed by neural models. The neural models use synapses or synaptic circuits to connect neurons to each other for exchanging signals. This invention consists of a new circuit composed of 2 transistors and one memristor (2T1R), serving as an artificial synapse for neuromorphic networks. The synapse is capable of communication and plastic modification (potentiation, depression) as a result of the timing between spikes delivered by neurons in the network. The synaptic plasticity allows to achieve learning and recognition of patterns (e.g. visual, auditory, etc.) in real time and with low power consumption.
Spike-timing dependent plasticity (STDP) characteristic has been successfully demonstrated for variable initial state and variable shape of the neuronal spikes. Unsupervised learning of a visual pattern has been demonstrated by simulations of a 2-layer neural network linked by 64 2T1R synapses (Fig. 1).
Campo di applicazione
The invention finds application in neuromorphic circuit for distributed systems of smart sensors. Although pattern learning and recognition is also possible via software, this invention permits to develop compact, low-power circuits for portable applications (e.g. cellular phones, smartwatches, vehicles, drones, etc.) and/or energy-autonomous (remote locations on the sea, in orbit, on buildings) for the unsupervised interaction with the real world (as opposed to, e.g., the web).
A neuromorphic system capable of real-time pattern recognition might find applications in monitoring of environment, population, public places and security.
The 2T1R synapse allows the scaling down of the sizes and complexity of the artificial synaptic circuit, thus obtaining a low power consumption device and achieving one of the important tasks in the design of the electronic neuromorphic network.
Stadio di sviluppo
Prototipo preliminare - Preliminary prototype.