Bio-inspired AI architectures

Neuromorphic engineering starts from an observation about cost. In a conventional accelerator, the arithmetic is cheap and moving the operands is expensive: fetching a weight from off-chip memory costs orders of magnitude more energy than the multiply it feeds. A clocked architecture pays this cost on every cycle for every unit, whether or not anything in the input has changed. Biological neurons do the opposite — they sit silent and consume almost nothing until an event arrives.

Where the energy actually goes

A spiking network encodes information in the timing of discrete events rather than in continuous activation values. Two consequences follow. First, a spike carries no magnitude, so the receiving neuron adds a stored weight to its membrane potential instead of multiplying: accumulate replaces multiply-accumulate. Second, and more important, computation happens only where and when a spike occurs. If a scene is mostly static, most of the array does nothing and draws only leakage. Memory is also placed next to the compute, so weights do not travel.

The saving is therefore conditional on sparsity, and this is the part usually glossed over. Feed a neuromorphic chip dense frames from an ordinary camera and the advantage largely evaporates, because the input must be converted into spike trains and the network becomes busy everywhere. The applications where the argument holds are those with a naturally event-driven front end — event cameras that report per-pixel brightness changes rather than frames, keyword spotting, vibration and radar monitoring — and with a power budget in the microwatt to milliwatt range where a conventional part cannot go. It is an edge-sensing argument, not a datacentre one.

Training has no biological implementation

Backpropagation requires each synapse to know weights it does not touch — the error signal must be routed backwards through the transpose of the forward weight matrix. No known mechanism transports weights that way in a brain, and this weight-transport problem, together with backpropagation’s need to hold every forward activation until the backward pass, is why it is not treated as a model of biological learning.

There is also a hard mathematical obstacle specific to spikes. A neuron fires when its membrane potential crosses a threshold, which is a step function; its derivative is zero everywhere and undefined at the threshold, so gradients vanish. Surrogate-gradient training accepts this and cheats deliberately: the forward pass uses the true threshold, while the backward pass substitutes the derivative of a smooth function — a sigmoid or a triangular window — in place of the step. It works well enough that most reported spiking-network accuracies come from it. What it is not is a learning rule a chip could run: training still happens offline in a conventional deep-learning framework on conventional hardware, and only the trained weights are deployed. On-chip local rules, spike-timing-dependent plasticity among them, remain far behind on supervised tasks.

The gap that keeps it in the lab

There is no shared software layer. Each family of parts — research chips such as Loihi and SpiNNaker, commercial edge devices from several vendors — carries its own toolchain, its own neuron model, its own quantisation and its own constraints on fan-out and per-core memory. A model is ported, not compiled. Benchmarks are not comparable across platforms, so an energy claim is rarely a like-for-like measurement. That absence of a common substrate, more than any device physics, is what keeps neuromorphic parts out of production systems.