Short- and long-term memory functions are mimicked through artificial synapse

Memory retention in the human brain is based on rehearsal and repetition. Inspired by biological neural networks and the functions of memory, Prasad et al. designed an artificial short- and long-term memory magnetic tunnel junction (SALT-MTJ) synapse.

“There have been past works which demonstrate short-term forgetting behavior in various electronic synaptic candidates, such as phase-change memories and magnetic tunnel junctions,” said author Nitin Prasad. “However, this is the first report which offers a way to tailor the short- and long-term memory lifetimes to suit the design requirements of neural networks.”

The authors developed this SALT-MTJ synapse by engineering a metastable state between two stable states of a magnetic tunnel junction stack. Whether or not the conductance of the synapse is switched between the two stable states is based on the input stimuli repetition rate encountered by the synapse relative to the lifetime of the metastable state. This conductance change relates to the switching between short- and long- term memory in neurological systems based on the type of stimulus being received.

This research was conducted in an effort to further neuromorphic computing – computing designs which mimic biological systems – while avoiding the von Neumann bottleneck, which is a limitation based on the inevitable delay in the transfer of data between the processor and the memory device.

“These artificial SALT-MTJ synapses could be used to realize stochastic binary neuromorphic networks to perform tasks such as image classification, speech recognition and more,” said Prasad.

The next step in this research will be the fabrication of the SALT-MTJ synapses to characterize the metastable short-term memory, and the theoretical estimation of the performance edge of these synapses in tasks such as digit recognition.

Source: “Realizing both short- and long-term memory within a single magnetic tunnel junction based synapse,” by Nitin Prasad, Tanmoy Pramanik, Sanjay K. Banerjee, and Leonard F. Register, Journal of Applied Physics (2020). The article can be accessed at https://doi.org/10.1063/1.5142418 .