Biological Inspiration: How Brains Compute
Neuromorphic Computing · Frontiers track
7 min read
In this module
The Brain as a Computer
Every computing paradigm we've studied so far — transistors, logic gates, CPUs, GPUs, even quantum computers — shares one design philosophy: they're engineered down from mathematical abstractions. Logic gates implement Boolean algebra. CPUs execute instructions from a formal ISA. Top-down designs, built from mathematical first principles.
Neuromorphic computing flips this. It looks at the most capable information-processing system we know of — the biological brain — and asks, "what if we built computers that work like this?"
The human brain packs roughly 86 billion neurons wired by something like 100 trillion synapses. It recognises faces in milliseconds, parses language in real time, and navigates cluttered environments — all on about 20 watts. The most powerful AI systems today burn megawatts to match some of that, and still fall short in ways that matter. Closing that efficiency gap is where neuromorphic computing starts.
How Biological Neurons Work
A biological neuron has three main components:
Dendrites - tree-like branches that receive input signals from other neurons. A single neuron may have thousands of dendritic connections.
Cell body (soma) - integrates all incoming signals. The soma maintains a membrane potential - a voltage difference between the inside and outside of the cell, typically around -70 millivolts at rest.
Axon - a long fibre that transmits output signals to other neurons. The axon branches at its end to form synaptic terminals that connect to the dendrites of downstream neurons.
Dendrites (inputs) ╲ │ ╱ ╲ │ ╱ ┌────────────┐ │ Cell Body │ ← Integrates inputs │ (Soma) │ └─────┬──────┘ │ Axon (output) │ ╱─────┼─────╲ ╱ │ ╲ Synaptic Terminals → connect to other neurons' dendritesThe Action Potential (Spike)
When the accumulated input raises the membrane potential above a threshold (approximately -55 millivolts), the neuron fires an action potential - a brief, stereotyped electrical pulse lasting about 1-2 milliseconds. This pulse, called a spike, travels down the axon to the synaptic terminals.
The key properties of a spike are:
- All-or-nothing: A neuron either fires or it does not. There are no "partial" spikes. This is fundamentally digital - a spike is a binary event.
- Stereotyped waveform: Every spike from a given neuron looks essentially the same. Information is not encoded in the shape or amplitude of the spike.
- Refractory period: After firing, a neuron cannot fire again for about 1-2 milliseconds (the absolute refractory period). This limits the maximum firing rate to about 500-1000 Hz.
The deepest difference between biological neural processing and digital computing is that spikes are events. They occur at irregular times set by the neuron's inputs, not on a fixed clock. A neuron with no input generates no spikes and burns almost no energy. A neuron driven hard might fire at 100 Hz. Compare that to a clocked digital circuit, where every gate flips every cycle whether or not it's doing anything useful.
Synapses: Where Learning Happens
The junction between two neurons is called a synapse. When a spike arrives at a presynaptic terminal, it triggers the release of neurotransmitter molecules that cross the synaptic cleft (a tiny gap of about 20 nanometers) and bind to receptors on the postsynaptic neuron's dendrite.
The effect on the postsynaptic neuron depends on the type of synapse:
- Excitatory synapses - increase the membrane potential, making the neuron more likely to fire. These use neurotransmitters like glutamate.
- Inhibitory synapses - decrease the membrane potential, making the neuron less likely to fire. These use neurotransmitters like GABA.
Each synapse has a weight - a strength that determines how much effect a presynaptic spike has on the postsynaptic neuron. Learning in the brain is primarily about adjusting these synaptic weights.
Synaptic Plasticity
Learning happens through synaptic plasticity — synapses that strengthen or weaken over time as activity flows through them. The headline rule is Hebbian learning, usually summarised as "neurons that fire together, wire together."
More precisely, Spike-Timing-Dependent Plasticity (STDP) adjusts synaptic strength based on the relative timing of pre- and postsynaptic spikes:
- If the presynaptic neuron fires just before the postsynaptic one (suggesting it helped cause the firing), the synapse is strengthened — Long-Term Potentiation, or LTP
- If it fires just after (suggesting it didn't contribute), the synapse is weakened — Long-Term Depression, LTD
The elegance is in the locality. Each synapse only needs to know its own pre- and postsynaptic spike times, not anything about the rest of the network. That's exactly the property you want when you're trying to put learning into silicon.
Brain vs. Computer: Key Differences
Why bother with neuromorphic computing at all? Because the brain and a conventional computer are radically different machines:
Clock Speed vs. Event-Driven
- Computer: Fixed clock (e.g., 3 GHz), all circuits switch every cycle
- Brain: No global clock; neurons fire asynchronously based on their inputs
Precision vs. Noise Tolerance
- Computer: 64-bit floating-point precision, exact deterministic results
- Brain: Noisy, imprecise signals; resilient to neuron death and synaptic variability
Serial Depth vs. Massive Parallelism
- Computer: Deep computational pipelines (billions of sequential operations per second on a few cores)
- Brain: Shallow processing depth (a visual recognition task involves perhaps 10-20 synaptic steps) but massive parallelism (billions of neurons operating simultaneously)
Separation vs. Co-location of Memory and Compute
- Computer: Von Neumann architecture - memory and processing are separate
- Brain: Memory (synaptic weights) and processing (neural integration and firing) are co-located at every synapse
Power Consumption
- Computer: A modern GPU consumes 300-700 watts; a data centre rack consumes 10-30 kilowatts
- Brain: ~20 watts for the entire organ
The brain's efficiency comes from a stack of design choices working together: event-driven computation (no energy on idle circuits), co-located memory and compute (no shuttling weights around), massive parallelism (billions of simple processors, not a handful of complex ones), and sloppy analogue maths that's just precise enough for the task. Neuromorphic computing tries to bottle as many of these tricks as possible into silicon.
Neural Coding: How Spikes Carry Information
If every spike looks the same, how do neurons encode different types of information? Several coding schemes have been identified:
Rate Coding
The simplest coding scheme: information is encoded in the firing rate (number of spikes per second). A brighter light causes retinal neurons to fire faster. This is analogous to a pulse-width modulation signal.
Temporal Coding
Information is encoded in the precise timing of individual spikes. A sound arriving at the left ear before the right ear creates a timing difference of microseconds, which the brain uses to localize the sound source. Temporal coding can carry more information per spike than rate coding.
Population Coding
Information is spread across a population of neurons. No single cell carries the whole signal; the pattern across many neurons does. You get redundancy for free (lose a neuron, the signal survives) and a high-dimensional representation in the bargain.
Burst Coding
Some neurons fire in rapid bursts (several spikes in quick succession) that carry different information than single spikes at the same average rate. Bursts may indicate particularly important or reliable signals.
From Biology to Silicon
Nobody's trying to clone a biological neuron exactly. Real neurons are messy, and we still don't understand half of what they do. What neuromorphic engineering extracts is the computational principles that make the brain efficient and rebuilds them in silicon: event-driven processing (compute only when spikes arrive), co-located memory and compute (synaptic weights live where they're used), massive parallelism across many simple processors, local learning rules that don't require a global error signal, and just enough analogue precision to get the job done.
In the next lesson, we examine the hardware that implements these principles: Intel Loihi, IBM TrueNorth, SpiNNaker, and other neuromorphic chips.