In the brain, information is transmitted through the release of chemical neurotransmitters at the junction, or synapse, between neurons. This process is stochastic in nature, which is believed to be a key feature in how the brain operates. To mimic the behaviour and performance of biological neurons, stochastic artificial neural networks — a particular class of machine learning algorithms — are being developed. However, increasingly complex operations require considerable computing power. To maximize energy efficiency, such networks must be ultimately translated into physical materials — that is, implemented in hardware. Now, writing in Nature Communications, Sourav Dutta and colleagues present a hardware implementation of a specific stochastic neural network, called a neural sampling machine (NSM).
The hardware implementation of the algorithm can be achieved through compute-in-memory (CIM) architectures, which, as the name suggests, allow joint data processing and storage in the computer memory. Non-volatile memory elements that retain information even when power is turned off are the main facilitators for CIM. “However, network connections in a traditional CIM architecture don’t exhibit the stochastic behaviour required for an NSM,” comments Dutta. The desired outcome can be implemented within existing CIM architectures by pairing the non-volatile memory element with a binary stochastic selector. For that purpose, the researchers used a ferroelectric field-effect transistor (FeFET) as the non-volatile memory element. The FeFET conductance is correlated to the synaptic weight and is modulated by an applied voltage. The stochastic selector is fabricated from a stack of Ag, TiN, HfO2 and Pt films that can switch between on and off states through the formation and the rupture of an Ag filament under external bias. “By stochastically turning on or off, the selector allows random access to the synaptic weight with a level of uncertainty. This proved to be a unique approach towards building scalable and energy-efficient hardware supporting NSM,” says Dutta.
This is a preview of subscription content, access via your institution