How it works
How neuromorphic computing works, shown live, and how it compares with the brain
A running simulation of an event camera feeding a small spiking network: watch neurons integrate, leak and fire, and see the work an event-driven chip does fall when nothing moves. Then a sourced comparison with the brain the idea comes from.
See it working
This is a model running in your browser, not an animation. A 24 × 16 event camera watches a bright spot, and 32 spiking neurons work out where the spot is from the events alone. Try the spot still, slow and fast, and add sensor noise.
Simulation time runs at a quarter of real time so single spikes are visible. Operation counts are per second of simulated time. Set the spot still: the sensor goes quiet and the event-driven work drops towards zero, while the clocked processor keeps reading every pixel. Turn on noise: the work rises and decoding gets worse, which is why real sensors filter noise events.
What you are looking at
The sensor works like the change-sensitive cells of a retina. Each pixel remembers the brightness it last reported and speaks only when that changes by more than a set contrast. A still scene produces almost nothing. A moving edge produces a burst of ON events where it arrives and OFF events where it leaves.
Each event is a small push into 32 neurons, strongest for the neurons tuned to that pixel's column. A neuron's potential leaks back towards zero, so a lone event fades. Several events close together push it over the threshold, and it fires a spike and resets. The pattern of spikes in the raster is the network's answer: which neurons are firing says where the spot is.
The counters compare two ways of doing the same job. A conventional processor reads every pixel at a fixed rate and multiplies all of them by all the weights, moving or not. An event-driven (neuromorphic) chip does work only when an event arrives: 32 additions per event here.
The neuron, in one equation
Symbols
- membrane potential (dimensionless here; millivolts in a real neuron)
- membrane time constant: how fast the potential leaks away
- synaptic weight from input i (here, how close pixel i's column is to the neuron's preferred position)
- the times of input i's events, and whether input i sent an event in step t
- firing threshold and refractory period
In plain words
This is the leaky integrate-and-fire neuron, the workhorse model of both computational neuroscience and neuromorphic chips. Between inputs the potential decays exponentially. Each incoming event adds its weight, and crossing the threshold makes a spike. The second line is exactly what the simulation computes every millisecond.
How it compares with the brain
| Human brain | Intel Hala Point (1,152 Loihi 2 chips) | This simulation | |
|---|---|---|---|
| Neurons | about 86 billion (Azevedo et al., 2009) | 1.15 billion | 32 |
| Signal | spikes about 1 ms long; analog, chemical synapses | digital spikes, binary or carrying a value | spikes in 1 ms steps |
| Typical activity | average cortical rate estimated from well under 1 Hz (Lennie, 2003) to about 4 Hz (Howarth et al., 2012) | event-driven: idle neurons cost little | a few spikes per neuron per second while the spot moves |
| Power | about 17 to 20 W from glucose (Balasubramanian, 2021) | up to 2,600 W | – |
| Power per neuron | about 0.2 nW | about 2.3 µW (at maximum draw) | – |
| Where the energy goes | communication about 35× computation in cortex (Levy and Calvert, 2021) | moving data costs more than arithmetic (Horowitz, 2014) | we count operations only |
The same ideas: information carried by spikes, potentials that leak, thresholds, very sparse activity, and memory (synapses) held right next to the units that use it.
The differences, in practice: the brain is analog, chemical and three-dimensional; its dendrites compute on their own, and it keeps learning all the time. Neuromorphic chips are digital and flat. They switch in microseconds rather than milliseconds, and most run fixed weights. They borrow the brain's operating principles, not its biology.
The lesson that matters for engineering
The most useful parallel is about cost, not neurons. Levy and Calvert estimate that the human cortex spends about 35 times more energy communicating than computing. Horowitz's figures say the same of chips: moving a number from memory costs far more than doing arithmetic with it. Neuromorphic designs keep memory beside each neuron and send something only when it changes. That is why they shine on sparse, event-like signals, and why our own studies focus there.
Sources
- Azevedo et al., Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain, J. Comp. Neurol. (2009)
- Lennie, The cost of cortical computation, Current Biology (2003)
- Howarth, Gleeson and Attwell, Updated energy budgets for neural computation in the neocortex and cerebellum, JCBFM (2012)
- Levy and Calvert, Communication consumes 35 times more energy than computation in the human cortex, PNAS (2021)
- Balasubramanian, Brain power, PNAS (2021)
- The Register, Intel Hala Point (April 2024)
- Gallego et al., Event-based vision: a survey (2022)
- Horowitz, ISSCC 2014