Glossary
The vocabulary of neuromorphic computing, in plain words.
- Neuromorphic
- Built on the organising principles of nervous systems. The word was introduced by Carver Mead.
- Spike
- A brief, all-or-nothing pulse by which neurons signal. Also called an action potential.
- Spiking neural network
- A network whose units communicate by spikes in time rather than by continuous numbers.
- Leaky integrate-and-fire
- The simplest spiking neuron model: sum the input, let it decay, fire at a threshold, reset.
- Synapse
- The connection through which one neuron influences another.
- Synaptic weight
- The strength of that influence.
- Plasticity
- Change in synaptic weights with experience; the basis of learning.
- STDP
- Spike-timing-dependent plasticity: a synapse strengthens if the sending neuron fires just before the receiving one, and weakens if just after.
- Rate coding and temporal coding
- Carrying information in how often a neuron fires, or in exactly when.
- Event
- A spike treated as a message: which unit fired, and when.
- Address-event representation
- A communication scheme in which each spike is sent as the address of its source.
- Event camera
- A vision sensor that outputs a stream of per-pixel brightness changes instead of frames.
- Asynchronous circuit
- A circuit with no global clock; parts act when their inputs change.
- Von Neumann bottleneck
- The limit set by moving data between a separate processor and memory.
- In-memory computing
- Performing arithmetic inside the memory array, where the data already are.
- Memristor
- A two-terminal element whose resistance depends on the current that has passed through it; usable as an analog synapse.
- Crossbar array
- A grid of wires with a memory element at each crossing; it performs a matrix–vector product in one step.
- Sparsity
- Most units being inactive at any moment. Event-driven hardware spends energy only on the active ones.
- Surrogate gradient
- A training method that substitutes a smooth function for the spike, so that spiking networks can be trained by gradient descent.
- Edge device
- A device that computes where the data are produced, on a small power budget.