Front. Neurosci. 10, 482 (2016)

The electrical resistance state of a memristor is a function of the history of currents and voltages previously applied to the device. Such behaviour closely resembles the activity-dependent plasticity of biological synapses, making memristors natural candidates to implement learning processes in neuromorphic hardware architectures. In analog memristors the possible resistance states are multi-valued, offering a closer analogy to biological synapses. However, the implementation of unsupervised learning processes based on networks of analog memristors has been rarely reported.

Now, Erika Covi and colleagues at the National Research Council in Italy and the University of Southampton in the UK have reported a network of 125 analog memristors capable of reproducing unsupervised learning tasks. Each memristor is a TiN/HfO2/TiN heterostructure and the researchers exploited filamentary resistive switching processes to mimic the synaptic plasticity therein. The memristor network connects two layers of artificial neurons — the so-called pre- and post-neurons. A 5 × 5 square array of pre-neurons, each of which act as a black or white pixel, provides images of letters as input signals to the network. The first activated post-neuron defines the output signal. After an unsupervised training process, the network was able to associate one specific post-neuron to each different letter being shown, demonstrating successful recognition. This was also verified for noisy or incomplete input images.