New Device Leads to “Dendrocentric Learning”

Neuromorphic computing has so far tried to mimic the synapses between neurons in the brain. But a new approach instead aims to act like dendrites, the spindly structures that branch out from the nucleus of a neuron like the roots of a tree. Dendrites receive signals from other neurons via synapses, transmitting them from tip to stem to the nucleus. In computing, “nanodendrites” could function similarly, according to a team of researchers at Stanford University.

Collaborating with semiconductor manufacturer GlobalFoundries, the researchers proposed one such nanodendrite at the 2023 IEEE International Electron Device Meeting (IEDM) this week. The device, a modified transistor, acts as a switch that detects a sequence of microsecond-long voltage pulses. It turns on, thus allowing current to pass, only if the pulses arrive in the correct order. According to Stanford bioengineering professor Kwabena Boahen, this approach could lead to efficient parallel processing in the 3D chips that AI will increasingly depend upon. By emulating the brain’s dendrites, these chips would use less energy and, importantly, generate less heat.

Heat presents a “fundamental issue” in today’s 3D chip technologies, says electrical engineer H.-S. Philip Wong, an IEEE Fellow and a professor of electrical engineering at Stanford. The heat generated grows in proportion to the volume—but the chips dissipate heat at a rate proportional to surface area. That’s why, currently, “all computational advances are limited by heat dissipation,” Wong says.

The problem can be solved by the nanodendrite approach, Wong suggests, because it uses voltage in discrete pulses instead of continuously held levels. It therefore activates fewer wires at any given moment and thus generates less heat.

A typical field-effect transistor consists of three terminals: the source, gate, and drain. For charge to move from the source to the drain, a voltage is applied to the gate, changing the electric field and the conductivity of the silicon. The Stanford device maintains the same basic elements, but it splits the transistor’s gate into three parts. It also embeds a thin layer of ferroelectric material in the multi-part gate, causing polarization to switch when an electric field is applied.