Glossary - Practical Spiking Neural Networks

Terms are listed alphabetically. Each entry gives the expansion of the term (where it has one) followed by a short definition. Reference an entry from any chapter with the {term} role, e.g. {term}`LIF` , so that the reader can jump straight here.

action potential: See spike.
AEDAT: Address Event Data — a family of file formats used to store recordings from event-based sensors such as a DVS.
AER: Address Event Representation — a format for communicating spikes as a stream of addresses, where each address identifies the neuron that fired and the event’s position in the stream carries its timing.
ALIF: Adaptive Leaky Integrate & Fire — a LIF neuron whose firing threshold rises after each spike and decays back over time, so that sustained input produces a falling firing rate.
ANN: Artificial Neural Network — a second-generation neural network whose neurons emit continuously valued activations rather than spikes, and which carries no state across time unless one is added explicitly.
BPTT: Backpropagation Through Time — training a stateful network by unrolling it across simulation time steps and backpropagating through the resulting graph, in which the weights are shared across every step.
CPU: Central Processing Unit — the general-purpose processor of a conventional computer, responsible for arithmetic, logic, control and I/O.
credit assignment: The problem of determining which neurons and synapses were responsible for an output error, and by how much each should change. In an SNN it must be solved in space and time.
CUBA: CUrrent-BAsed — a neuron model in which an incoming spike injects a current that then charges the membrane, as opposed to a conductance-based (COBA) model, in which the spike changes the membrane’s conductance.
decoding: Converting spike trains back into continuous values, so that the output of a spiking network can be interpreted. See also rate decoding and spike trace.
DSP: Digital Signal Processor — a processor specialised for numerical operations on sampled signals, often present alongside a neuromorphic accelerator to carry out preprocessing.
DVS: Dynamic Vision Sensor — a camera whose pixels asynchronously emit events when the luminance they observe changes, rather than reporting whole frames at a fixed rate.
eligibility trace: A per-synapse record of recent pre- and post-synaptic activity that marks the synapse as eligible for a weight update when a learning signal later arrives, allowing credit to be assigned backwards in time without storing the full history.
encoder: A spike generator designed to accept an input signal and represent it as spikes, i.e. to represent a dense continuous-valued signal as a sparse sequence of discrete events.
encoding: Converting continuous-valued data into spikes so that it can be fed to a spiking network. Broadly divided into rate encoding and temporal encoding.
GPU: Graphics Processing Unit — a highly parallel accelerator, originally built for graphics, now the standard hardware for training neural networks.
graded spike: A spike carrying an integer amplitude greater than one, rather than the binary value used by most models. Biological action potentials have no equivalent notion of amplitude.
GRF: Gaussian Receptive Field — a population encoding scheme in which each neuron responds to a Gaussian-shaped window of the input range, and the distance from an input to a neuron’s centre sets that neuron’s spike latency.
hard reset: Setting the membrane potential to $V_\text{rest}$ after a spike, discarding any charge accumulated beyond the threshold. Contrast soft reset.
Heaviside step function: The function $\Theta$ that returns 1 for a positive argument and 0 otherwise, used to convert a membrane potential crossing into a spike. Its derivative is zero almost everywhere, which is why SNNs need surrogate gradients.
Hodgkin-Huxley: A spatial neuron model that reproduces the generation of an action potential from the dynamics of individual ionic conductances, at considerably greater computational cost than a point neuron.
IF: Integrate & Fire — the simplest point neuron model, which integrates input into a non-decaying membrane potential and spikes when that potential crosses the threshold. Contrast LIF.
LIF: Leaky Integrate & Fire — a point neuron model that integrates input into a decaying membrane potential, so that charge accumulated in the absence of further input leaks away over time. The most widely used neuron model in SNNs.
LTD: Long-Term Depression — a persistent decrease in synaptic strength, typically produced when the pre-synaptic neuron fires after the post-synaptic one. See STDP.
LTP: Long-Term Potentiation — a persistent increase in synaptic strength, typically produced when the pre-synaptic neuron fires shortly before the post-synaptic one. See STDP.
membrane potential: The internal state variable $V$ of a spiking neuron, which accumulates incoming current and triggers a spike on crossing the threshold. Also called the neuron’s voltage.
mismatch: Device-to-device variation introduced by the manufacturing process, which causes nominally identical analog or mixed-signal neurons on the same chip to behave differently. Models targeting such substrates must be trained or validated to tolerate it.
NAS: Neural Architecture Search — automated search over the space of possible network topologies for one that best meets a target objective.
NEF: Neural Engineering Framework — a framework for building spiking networks that compute specified functions, grounded in representation, transformation and dynamics principles.
NoC: Network-on-Chip — a communication architecture that interconnects multiple processing elements on a single chip.
ODE: Ordinary Differential Equation — an equation describing how a function changes with respect to one or more variables. The continuous-time equations of the IF and LIF neurons are first-order ODEs.
plasticity: The capacity of a synapse to change its strength in response to neural activity, and hence the basis of learning in a spiking network.
point neuron: A neuron model that ignores the spatial structure of a biological neuron — its ionic channels, axial conductances and axonal propagation — and reduces it to a single state variable. Contrast spatial neuron.
population coding: Representing one input dimension with a group of differently tuned neurons rather than a single neuron, so that distinct characteristics of the signal can be captured separately.
quantization: Reducing the numerical precision used to represent weights, neuron state and activations, so that a model fits the arithmetic and memory available on the target hardware.
R&F: Resonate & Fire — a point neuron model whose sub-threshold dynamics oscillate, making it selectively responsive to inputs near its resonant frequency.
rank order coding: A temporal scheme that carries information in the order in which neurons fire rather than in their precise spike times. Commonly abbreviated ROC, but note that in machine learning ROC almost always denotes the receiver operating characteristic; this book prefers the full name to avoid the clash.
rate decoding: Recovering a value from a spike train by averaging its spikes over the simulation window, so that a higher spike count indicates stronger evidence.
rate encoding: Representing a continuous value as a spike rate, such that the number of spikes per unit time is proportional to the value being encoded.
refractory period: A short interval after a spike during which a neuron is prevented from firing again, mimicking the biological neuron’s reluctance to produce two action potentials in immediate succession.
ReLU: Rectified Linear Unit — the activation function $\max(x, 0)$ , widely used in ANNs. A LIF neuron’s firing-rate curve can be tuned to closely approximate it.
SNN: Spiking Neural Network — a third-generation neural network in which neurons carry state across time and communicate through discrete spikes rather than continuous activations.
soft reset: Subtracting the threshold $V_\text{thr}$ from the membrane potential after a spike, so that charge accumulated beyond the threshold is retained rather than discarded. Contrast hard reset.
spatial neuron: A neuron model that represents the physical extent of a biological neuron, including axonal propagation delays and dendritic structure. Contrast point neuron.
spike: A discrete, all-or-nothing event emitted by a neuron when its membrane potential crosses threshold; the unit of communication in an SNN. Also called an action potential.
spike generator: A programming construct that produces a spike train according to some chosen implementation. It does not necessarily take an input; one that does is an encoder.
spike trace: A smoothed, continuous-valued signal obtained by low-pass filtering a spike train, which tracks the neuron’s mean firing rate. Also called synaptic filtering.
spike train: A sequence of spikes ordered in time, usually represented as a vector whose index denotes the time step and whose elements are 0 or an integer.
SRM: Spike Response Model — a formulation that decomposes a neuron into linear filters followed by a single threshold nonlinearity, which is what makes surrogate gradients principled: only that one nonlinearity needs approximating.
STDP: Spike-Timing-Dependent Plasticity — a learning rule that adjusts a synaptic weight according to the relative timing of the pre- and post-synaptic spikes, combining LTP and LTD.
substrate: The physical medium from which the neurons and synapses of a device are built — analog, digital or mixed-signal. The choice fixes the numerical precision the hardware can hold, the effort of porting to a newer process node, and the maturity of the surrounding tools, and it is among the hardest design decisions to undo. See also mismatch.
surrogate gradient: A smooth function substituted for the derivative of the Heaviside step function during the backward pass, leaving the forward pass discrete. The standard method for training SNNs with gradient descent.
temporal encoding: Representing a value in the timing of spikes rather than in their rate, which avoids the integration window that rate schemes require. See TTFS and GRF.
TinyML: Machine learning deployed on severely resource-constrained embedded devices, where memory, latency and energy budgets dominate every design decision.
TTFS: Time-to-First-Spike — a latency encoding scheme in which each neuron fires at most once, with a delay inversely related to the magnitude of its input, so that stronger stimuli spike sooner.
WTA: Winner-Takes-All — a mechanism in which the most strongly stimulated neuron in a group suppresses the activity of the others.