A finite system cannot give every input an unlimited response. It has to scale the world to fit the machinery it has.
Walk into a dim room.
A small difference in light matters.
Walk into bright daylight.
The same absolute difference may matter far less.
A useful sensory system therefore cannot respond only to raw magnitude.
It often needs context.
Normalisation is one way to build that context into the response.
Quick Route
- Gain: sensitivity of output to input.
- Dynamic range: the useful span between detection and saturation.
- Normalisation: response scaled relative to a contextual pool or reference level.
- Adaptation: change in coding properties over time in response to stimulus statistics or state.
- Compression: broader operation mapping a large range into a smaller representation.
Canonical Job
Normalisation owns one reader job in Cognitive Art:
How can a system rescale one response according to the activity or magnitude of a broader context so finite output range remains useful?
The canonical neural version is divisive normalisation.
In simplified form, a neuron’s driven response is divided by a term that depends on activity in a normalization pool.
The landmark Nature Reviews Neuroscience review Normalization as a Canonical Neural Computation describes this operation across visual, olfactory, auditory, multisensory, attentional and value-related systems and argues that similar computations can be implemented by different circuits.
One-sentence answer: Normalisation is a context-dependent rescaling operation in which a response is adjusted relative to a broader pool or reference signal, helping finite systems preserve useful sensitivity across changing input conditions.
Normalisation Is Not Statistical Normalisation
Data science uses “normalization” for many preprocessing operations.
- rescale values to 0–1,
- subtract a mean,
- divide by standard deviation,
- normalise vector length.
Those are legitimate mathematical operations.
They are not automatically divisive neural normalisation.
This article owns the neural and general systems idea of context-dependent response rescaling, not every data-preprocessing convention that shares the word.
Normalisation Is Not Unicode Normalization
Encoding systems use “normalization” for transforming equivalent text representations into standard forms.
That is a separate owner in the encoding estate.
It has nothing to do with neural gain control.
The identical word should not collapse distinct reader jobs.
Normalisation Is Not Gain
The new Cognitive Art article What Is Gain? owns sensitivity of the input–output relation.
Normalisation is one operation that can alter effective gain.
Gain says:
how steep is the response?
Normalisation says:
what contextual quantity rescales that response?
Normalisation Is Not Adaptation
Adaptation describes change in sensitivity or coding properties over time after the environment changes.
Normalisation can operate rapidly and instantaneously relative to current population activity.
The two can interact.
They are not identical mechanisms.
Normalisation Is Not Compression
The existing How Compression Works article owns the broad job of representing information with fewer resources or a smaller range.
Normalisation is one specific rescaling computation.
It often compresses responses.
But not every compression method is normalisation.
The Divisive Form
A simplified normalisation model has a numerator carrying the neuron’s drive and a denominator carrying a semi-saturation constant plus pooled activity.
Schematic form:
response = drive / (constant + pooled activity).
As contextual activity grows, the same drive produces a smaller normalised response.
The computation expresses activity relative to context.
The Normalisation Pool
Which neighbouring or related signals contribute to the denominator?
That set is the normalisation pool.
The pool may depend on:
- spatial location,
- feature tuning,
- sensory modality,
- network connectivity,
- task or attention.
Pool definition is not a minor detail.
Change the pool and the contextual scaling changes.
Normalisation and Dynamic Range
The new Cognitive Art article What Is Dynamic Range? owns the usable input window.
Normalisation helps shift sensitivity toward differences relevant to the current context.
Carandini and Heeger explicitly identify maximising sensitivity and widening effective dynamic range among the computational benefits of normalisation.
This is why a finite firing-rate range can remain informative across different overall contrast or intensity levels.
Absolute Value Becomes Relative Value
Raw response:
how much input is present?
Normalised response:
how large is this drive relative to the surrounding pool?
This relative coding can support invariance.
An object’s identity can remain discriminable even as overall contrast or concentration changes.
Normalisation and Invariance
The Cognitive Art article What Is Invariance? owns what remains stable under transformation.
Normalisation can help create invariance to nuisance dimensions such as overall intensity.
For example, the visual system needs to represent pattern while contrast changes.
Rescaling relative to contrast context can reduce dependence on raw magnitude.
Normalisation and Contrast
Divisive normalization was developed in part to explain nonlinear response properties in primary visual cortex.
Increase contrast.
Neural responses rise but often saturate.
Add surrounding or cross-orientation stimulation.
Responses can be suppressed relative to what simple linear summation predicts.
The normalisation model captures many such effects through pooled divisive suppression.
Normalisation and Attention
Attention can change neural drive and/or the effective normalisation pool.
The normalization model of attention explains why attentional effects can look like response gain in some experiments and contrast gain in others.
This is a good example of why gain and normalisation need separate owners.
Observed slope change is one level.
Context-dependent divisive computation is another.
Normalisation and Multisensory Integration
Normalisation models have also been applied to multisensory responses.
Visual and vestibular signals, for example, can be combined in ways where pooled activity shapes response magnitude.
But not every multisensory computation must use one universal divisive circuit.
A 2025 Journal of Neuroscience paper, Beyond Divisive Normalization: Scalable Feedforward Networks for Multisensory Integration Across Reference Frames, explicitly explores alternative scalable mechanisms for some multisensory computations.
This is exactly why a canonical computation should not be turned into a universal monopoly.
Normalisation and Value
Relative value is often context-dependent.
Normalisation models have been proposed for value representation and decision systems where response to one option depends partly on the set of alternatives.
This does not mean every context effect in choice is divisive normalisation.
It means relative coding is one candidate computation that can be tested quantitatively.
Different Circuits Can Implement Similar Computation
One of the most important points in the canonical normalisation review is that the computation may recur without one universal biological mechanism.
Proposed mechanisms include:
- presynaptic inhibition,
- shunting inhibition,
- synaptic depression,
- balanced amplification,
- recurrent network interactions.
Same computational form.
Different implementation.
This is a central Cognitive Art lesson about levels of explanation.
Current Work: Recurrent Normalisation
Recent modelling continues to investigate how recurrent circuits can implement divisive normalisation stably.
The 2024 NeurIPS paper Unconditional Stability of a Recurrent Neural Circuit Implementing Divisive Normalization provides a formal recurrent implementation and analyses its stability.
A related 2025 paper, Stabilization of Recurrent Neural Networks Through Divisive Normalization, examines how divisive normalisation can stabilise recurrent dynamics.
These are computational models.
They strengthen understanding of what normalisation can do without proving that every biological circuit uses that exact implementation.
Current Work: Normalisation and Coding Theories
A 2025 PLOS Computational Biology paper, Relating Sparse and Predictive Coding to Divisive Normalization, explores mathematical relationships among divisive normalisation, sparse coding and predictive coding.
The careful conclusion is not that these theories are identical.
It is that several efficient-coding objectives can produce related computations under particular assumptions.
Normalisation Is Contextual Competition
When the normalisation pool becomes stronger, one response can shrink even if its direct input stays constant.
This creates a form of competition for finite response range.
But “competition” is an interpretation.
The measurable operation is rescaling by pooled activity.
Normalisation and Reference Dependence
The same numerator can produce different output because the denominator changes.
This is the mathematical heart of context dependence.
The representation stops asking only:
how big is this signal?
It asks:
how big is this signal compared with the activity that defines its current context?
Normalisation in Machine Learning Is Not One Thing
Artificial neural networks use batch normalization, layer normalization, group normalization and other techniques.
These methods often stabilise optimisation or control activation statistics.
They are not biologically identical to divisive normalization.
A neuroscience primer comparing artificial and biological networks notes that machine-learning normalization often subtracts means and divides by estimated standard deviations, whereas biological divisive normalization uses a different pooled-response computation.
Shared word.
Different algorithm.
Normalisation in Education: Use as Analogy Only
A score of 80 can mean something different on an easy paper and a difficult paper.
We sometimes rescale marks relative to a cohort or task distribution.
That is statistical normalisation or standardisation, not neural divisive normalisation.
The conceptual parallel is limited but useful:
raw magnitude can become more interpretable when placed relative to a context.
Normalisation in Organisations
A team produces 100 cases.
Is that high?
Relative to what capacity, staffing and case difficulty?
Organisations often construct normalised metrics to make context visible.
Again, these are statistical or managerial normalisations.
The neural article’s deeper contribution is the systems principle that finite response needs context-sensitive scaling.
Failure 1: Every Rescaling Is Divisive Normalisation
A value is standardised and the neural term is borrowed.
Repair: define numerator, denominator and normalization pool explicitly.
Failure 2: Gain Equals Normalisation
A slope change is measured and divisive normalisation is assumed.
Repair: establish the contextual pooling computation separately.
Failure 3: Canonical Means Universal Mechanism
Because normalisation appears across systems, one circuit implementation is assumed everywhere.
Repair: separate computation from implementation.
Failure 4: Normalisation Is Always Beneficial
Contextual suppression is assumed to improve every representation.
Repair: test which dimensions gain invariance and which information gets discarded.
Failure 5: Context Effect Proves Normalisation
An output changes with surrounding stimuli and divisive normalisation is declared the cause.
Repair: compare alternative mechanisms quantitatively and perturb the candidate pool.
Repair Path
- Define the direct drive.
- Define the candidate normalisation pool.
- Measure response across pool strength.
- Fit divisive and alternative models.
- Test effects on gain and saturation.
- Measure invariance and information loss.
- Perturb candidate circuit mechanisms where possible.
- Keep computation and implementation as separate claims.
The Normalisation Audit
- What response is being normalised?
- What provides the direct drive?
- What defines the normalization pool?
- Is the operation genuinely divisive?
- How does gain change?
- How does saturation change?
- Does effective dynamic range improve?
- Which invariances are gained?
- Which information is discarded?
- What circuit evidence distinguishes normalization from alternatives?
Research Notes and Further Reading
The central review remains Carandini and Heeger, Normalization as a Canonical Neural Computation (Nature Reviews Neuroscience).
For the relation between normalisation and cortical gain modulation, see Ferguson and Cardin, Mechanisms Underlying Gain Modulation in the Cortex.
For recent recurrent-circuit theory, see Unconditional Stability of a Recurrent Neural Circuit Implementing Divisive Normalization (2024) and Stabilization of Recurrent Neural Networks Through Divisive Normalization (2025).
For current work relating normalisation to broader efficient-coding theories, see Relating Sparse and Predictive Coding to Divisive Normalization (PLOS Computational Biology, 2025).
World Return
A normalisation model earns trust when changing the proposed pool changes response gain, saturation or invariance in the quantitative way the model predicts.
If the context effect survives after the supposed denominator is removed, the model needs another mechanism.
Final Thought: Context Can Be Part of the Measurement
A response does not always mean:
how much input arrived?
Sometimes it means:
how large was this input compared with everything else currently competing for the same finite response range?
That is why normalisation is more than scaling.
It is a way of making context part of the code.