Technology
The Role of Memristors in Neuromorphic Computing
Quick fact
The memristor was first theorized in 1971 by Leon Chua, but it was not physically realized until 2008 by HP Labs, which ignited the modern field of memristor-based neuromorphic computing.
Why this is interesting
Your brain performs massively parallel, energy-efficient computations, yet it has no separate memory and processor. How can we build a computer that works like that?
Read the full explanation
Understanding The Role of Memristors in Neuromorphic Computing
Imagine a light switch that not only turns lights on and off but also remembers the last position even when the power is off. A memristor is a two-terminal electronic component whose resistance can be changed by applying voltage, and this resistance remains unchanged when power is removed. This 'memory resistance' property makes it an ideal analogue for biological synapses, which change their strength based on activity. In neuromorphic computing, we build circuits that mimic brain neurons and synapses. A neuron accumulates signals; when its threshold is reached, it fires an output. The synapse between neurons modulates the signal's strength. A memristor placed between neurons can serve as that synapse: the conductance (inverse of resistance) represents the synaptic weight. Learning involves adjusting that weight, and because memristors can be written and read electrically, we can implement learning rules such as spike-timing-dependent plasticity directly in hardware.
A deeper explanation
The underlying mechanism of a memristor involves ion migration within a thin film of material (e.g., titanium dioxide). Applying a voltage moves oxygen vacancies, changing the conductivity of the film. This change is non-volatile and can be precisely modulated, allowing for continuous weight values. In a neural network, the multiplication of input signal and weight is performed by Ohm's law (current = voltage × conductance) on the memristor, and the summing of many such signals occurs naturally via Kirchhoff's current law. This means that matrix-vector multiplication, the core operation of neural networks, occurs physically in a crossbar array of memristors in a single time step, without moving data between memory and processor. This in-memory and analog computation vastly reduces energy consumption compared to traditional digital computers, which bus data back and forth. For neuromorphic systems, memristors enable scalable, low-power hardware that can learn and recognize patterns in real-time, advancing applications in edge AI, robotics, and brain-machine interfaces.