Foundations

How brain-inspired computing actually works

Start with a cell, end with a chip. Each idea below builds on the last, and none of it assumes a neuroscience or hardware background.

01 · Definition

What is neuromorphic computing?

Neuromorphic computing is engineering that borrows the brain's structure rather than just its vocabulary. Instead of running a neural network as software on a general-purpose processor, the hardware itself is shaped like a network.

Event-driven

Computation happens only when a signal arrives. Between events the circuit is effectively asleep, which is why power draw tracks activity rather than time.

Co-located memory

Synaptic weights are stored inside the core that uses them. There is no bus trip, so the dominant energy cost of conventional computing simply disappears.

Massively parallel

Every neuron runs independently and asynchronously. Scaling means adding more neurons, not raising a clock frequency that no longer wants to rise.

In plain terms: a conventional computer is a very fast clerk working through a list of instructions one at a time. A neuromorphic computer is a crowd of very simple clerks who mostly sit still and shout to each other only when something changes. For adding up a spreadsheet the first is far better. For noticing that something just moved in the dark, the second wins by orders of magnitude.

02 · Biology

The biological neuron

Roughly 86 billion of these cells, each connected to thousands of others, produce everything you have ever thought — on the power budget of a dim light bulb.

The four-step cycle

1ReceiveDendrites collect chemical signals from thousands of upstream neurons, each arriving with a different strength.
2IntegrateThe soma sums those inputs as an electrical charge on its membrane, while some charge continuously leaks away.
3FireCross about −55 millivolts and the cell emits an action potential: a millisecond-long pulse of fixed size.
4ResetThe membrane drops back and enters a brief refractory period during which no further spike is possible.

Biology mapped to silicon

Biology	Hardware
Dendrite	Input fan-in / router
Soma membrane	Capacitor or digital accumulator
Threshold	Comparator
Action potential	Address-event packet
Synaptic weight	Memristor conductance or SRAM value
Plasticity	Local STDP learning engine
Refractory period	Core-level timer

03 · Algorithms

Spiking neural networks

Often called the third generation of neural networks. The first used binary threshold units, the second continuous activations, and the third adds the one dimension the others discarded: time.

Coding schemes

How spikes carry meaning

Rate coding: The value is how often a neuron fires. Simple and robust, but it needs many spikes and therefore time and energy.
Temporal coding: The value is when the neuron fires. A single well-timed spike can carry information, which is extremely efficient.
Population coding: The value is which neurons in a group fire together, spreading one quantity across many cells for noise tolerance.

The model

Leaky integrate-and-fire

v ← v · (1 − leak) + Σ wᵢ · sᵢ(t)

if v ≥ threshold:
    emit spike
    v ← 0
    refractory ← 3 ms

Five lines describe the workhorse neuron of computational neuroscience — and it is exactly the model running in this site's simulator. More faithful models such as Izhikevich or Hodgkin–Huxley add richer dynamics at greater cost.

Learning

STDP and surrogate gradients

Spike-timing-dependent plasticity strengthens a synapse when the input fires just before the output and weakens it when the order reverses — cause before effect, measured in milliseconds. It is local, unsupervised and cheap enough to run on-chip.

For supervised accuracy, researchers instead train with surrogate gradients, which replace the spike's discontinuous step with a smooth stand-in so backpropagation can flow, or convert a conventional network into a spiking one after training.

04 · Devices

Memristors: the synapse in one component

A memristor's resistance depends on the current that has already passed through it. That single property makes it a physical memory of its own history — which is precisely what a synaptic weight is.

A crossbar array: apply voltages on the rows, read summed currents on the columns.

Analog by nature

Conductance varies continuously, so one device stores a graded weight instead of a single bit.

Non-volatile

The state survives power loss, enabling sensors that sleep for months and wake up already trained.

Physics does the maths

Ohm's law multiplies and Kirchhoff's law adds, so a crossbar performs an entire matrix multiplication in one electrical step.

The hard part

Devices drift, differ from one another and eventually wear out. Taming that variability is the field's central materials challenge.

05 · Silicon

The landmark chips

Two research programmes defined what a neuromorphic processor looks like. One proved the efficiency case; the other made the network able to learn while it runs.

IBM TrueNorth

2014

Neurons: 1,000,000
Synapses: 256,000,000
Cores: 4,096
Power: ~70 mW
Learning: Off-chip only

TrueNorth proved the thesis. A fully digital, entirely event-driven chip ran real-time vision at a power level a wristwatch could supply. Its neurons are fixed at design time — you compile a trained network onto it rather than teaching it in place — but as a demonstration that a million neurons could live inside a milliwatt budget, it reset expectations for the whole field.

Intel Loihi 2

2021

Neurons: ~1,000,000 / chip
Synapses: 120,000,000
Cores: 128 neuromorphic
Process: Intel 4
Learning: Programmable on-chip

Loihi's distinguishing feature is plasticity. Each core contains a small learning engine that can run spike-timing-dependent rules while the chip is operating, so a network can adapt to a new smell, surface or speaker without retraining in a data centre. Loihi 2 adds graded spikes — pulses that carry a small payload — and is programmed through the open-source Lava framework.

Also worth knowing

SpiNNaker

Manchester's million-ARM-core machine, built to simulate biological networks in biological real time.

BrainScaleS

Heidelberg's analog system runs up to ten thousand times faster than biology, compressing a day of learning into seconds.

Akida & Speck

Commercial edge parts from BrainChip and SynSense bringing event-driven inference to consumer devices.