A neural population can contain information without that information ever changing what the brain does next.
This is the readout problem.
Suppose one hundred neurons change their activity while an animal decides whether to turn left or right.
A machine-learning decoder can look at those neurons and predict the upcoming choice with high accuracy.
Tempting conclusion:
the brain reads those neurons in the same way to make the choice.
That conclusion is stronger than the evidence.
The external decoder has proved that choice-related information is available in the recorded activity.
It has not yet proved which downstream circuit receives that activity, what weights it applies, what temporal window it uses, whether the transformation is linear, whether the signal arrives before the behavioural deadline, or whether perturbing that candidate readout changes the decision.
A neural readout is where representation becomes consequence.
Quick Read
- Population code: information carried by a pattern across many neurons.
- Decoder: an analysis rule that estimates some variable from recorded neural activity.
- Neural readout: the biological transformation by which downstream activity is influenced by an upstream population.
- Readout weights: how strongly different neural dimensions or neurons contribute to a particular output.
- Output-potent dimension: a direction in population activity space that changes the selected readout.
- Output-null dimension: a direction in population activity space that can vary without changing that selected readout.
- Functionality: evidence that neural activity is actually used downstream, not merely correlated with or informative about a feature.
One-sentence answer: A neural readout is the downstream biological operation that converts activity in an upstream neural population into another neural signal, decision, command or action.
Canonical Job
Neural Readout owns one reader job in Cognitive Art:
How does information that exists in a neural population become a signal that a downstream circuit actually uses?
The Cognitive Art article What Is a Population Code? owns the distributed representation.
Neural Readout owns the transformation from that representation into downstream consequence.
This distinction is strongly supported by current conceptual work. The 2026 Nature Reviews Neuroscience perspective Clarifying the Conceptual Dimensions of Representation in Neuroscience distinguishes several kinds of evidence for representation, including sensitivity, specificity, invariance and functionality—whether the neural response is used downstream in the brain.
That final dimension is the readout question.
Readout Is Not Decoding
This is the most important boundary in the article.
A decoder is something an investigator builds.
A readout is something the biological system does.
External decoding asks:
given these recorded neural responses, can an algorithm infer variable X?
Biological readout asks:
does some downstream circuit receive and use these response dimensions in a way that changes its state, computation or behaviour?
A successful decoder proves statistical availability.
It does not automatically prove biological use.
The Decoder Can Know More Than the Brain Uses
Record ten thousand neurons.
Train a powerful nonlinear classifier with millions of parameters.
The classifier extracts a subtle feature with 95% accuracy.
Could the downstream biological circuit do the same?
Maybe.
But the external decoder may have advantages the brain does not:
- access to neurons that do not project to the same target,
- access to activity after the behavioural decision has already occurred,
- arbitrarily complex nonlinear computation,
- training labels unavailable to the organism,
- many trials for optimisation,
- offline computation unconstrained by biological latency.
Decoder success is therefore evidence about information in the recording.
It is not a licence to invent a biological readout.
Readout Is Not Population Code
A population code can exist before we know how it is read.
Several neural dimensions may encode:
- stimulus identity,
- context,
- confidence,
- movement direction,
- time in trial.
One downstream target may care mainly about movement direction.
Another may care about confidence.
A third may receive the same source neurons but weight them differently.
Population Code asks what structure is present.
Readout asks what part of that structure gets transmitted into this particular consequence.
Readout Is Target-Specific
A cortical area rarely has only one downstream destination.
Different neurons can project to different targets.
Different targets can therefore receive different slices of the same local population.
A 2025 Nature Neuroscience study, Specialized Structure of Neural Population Codes in Parietal Cortex Outputs, examined populations of cortical neurons projecting to specific downstream targets and reported target-specific correlation structure that enhanced population-level information relevant to accurate behaviour.
This supports a crucial refinement:
the biologically relevant population code may be defined partly by who receives it.
A code measured across every recorded neuron is not necessarily the code available to one downstream structure.
The Linear Readout
The simplest readout model is a weighted sum.
Let the population activity be a vector:
r = [r₁, r₂, r₃, …, rₙ].
A linear readout applies weights:
z = bᵀr.
Each weight says how strongly one neural activity dimension contributes to output z.
The output might represent:
- a motor command,
- a decision variable,
- input to another cortical population,
- activation of a downstream nucleus.
This model is powerful because a synaptic projection naturally resembles weighted summation.
It is also an approximation.
Linear Readout Is Not a Universal Biological Law
Downstream neurons are nonlinear biological systems.
They contain:
- thresholds,
- saturation,
- dendritic nonlinearities,
- short-term synaptic dynamics,
- inhibition,
- recurrent feedback,
- state dependence.
The Cognitive Art article What Is Nonlinearity? owns this broader departure from simple addition.
A linear readout is useful when it predicts the downstream transformation well enough for the task.
If the output depends on interactions that the weighted sum misses, the readout model must become nonlinear.
Mixed Selectivity Can Make Linear Readout Powerful
The Cognitive Art article What Is Mixed Selectivity? owns nonlinear combinations of task variables within individual neural responses.
This creates an elegant division of labour.
Upstream nonlinear mixed selectivity can expand population activity into a higher-dimensional space.
Once the relevant task conditions become separated in that space, a simple downstream linear readout can extract the desired combination.
Complex representation upstream.
Simple selection downstream.
This is one reason linear separability belongs inside the readout story rather than needing a separate Cognitive Art URL.
Output-Potent Dimensions
Now imagine the population as a high-dimensional space.
Move the population state in one direction.
The selected downstream output changes.
That direction is output-potent for that readout.
The 2024 Nature Reviews Neuroscience review Preparatory Activity and the Expansive Null-Space formalises this language for motor-cortical population dynamics.
For a linear readout z = bᵀr, a neural dimension v is output-potent if moving along v changes z.
In geometric terms:
v has some component aligned with the readout dimension b.
Output-Null Dimensions
Now move population activity in a direction orthogonal to that readout.
The neural population changes.
The selected output does not.
That direction is output-null for that readout.
This is one of the most powerful ideas in modern population neuroscience because it explains how a network can change internally without prematurely changing an external action.
Preparation Without Movement
Before a movement begins, motor cortex can already become highly active.
Why does the body not move immediately?
The classic 2014 Nature Neuroscience study Cortical Activity in the Null Space: Permitting Preparation Without Movement proposed and tested an output-null explanation.
Preparatory activity occupied dimensions that changed cortical state while largely cancelling at the level of the population readout driving movement.
Then, during execution, activity entered output-potent dimensions.
This turns a simple puzzle into geometry:
the brain can move internally without moving externally if the movement occurs in dimensions the current output cannot see.
Null Does Not Mean Useless
The word “null” sounds like nothing.
That is misleading.
Output-null activity can still participate in:
- preparation,
- internal computation,
- memory,
- state control,
- feedback correction,
- setting initial conditions for later potent activity.
Null means only:
this dimension does not affect this particular readout under this particular readout model.
Change the target or the readout weights and yesterday’s null dimension can become potent.
There Is No Universal Null Space
A dimension can be null for muscle output and potent for another cortical area.
It can be null for one choice and potent for another.
It can be null under one network state and potent after synaptic weights change.
Output-null is therefore relational:
null with respect to which receiver?
This is a direct Cognitive Art connection to Perspective and Context.
The Same Population Can Support Several Readouts
Suppose population activity contains three useful axes:
- stimulus identity,
- confidence,
- movement direction.
One downstream circuit weights the stimulus axis.
Another weights confidence.
A motor pathway weights movement direction.
The same upstream population can therefore be reused.
Different receivers extract different functions.
This is one reason population coding is more flexible than a one-neuron–one-message picture.
Readout as a Change of Coordinates
High-dimensional population activity is difficult to interpret neuron by neuron.
A readout chooses one direction or subspace that matters to the receiver.
This is mathematically similar to a projection.
But the Cognitive Art article What Is Projection? retains the general concept.
Neural Readout owns the biological consequence:
which projection does the receiving circuit actually implement?
Readout and Manifold
The Cognitive Art article What Is a Manifold? owns lower-dimensional structure embedded in high-dimensional state space.
A neural population may occupy a manifold.
A readout may depend on only a few directions within that manifold.
Recent population neuroscience increasingly focuses on such low-dimensional structures rather than individual-neuron labels. A 2025 Nature Reviews Neuroscience piece, Neural Manifolds: More Than the Sum of Their Neurons, describes the field’s shift toward coordinated population activity and latent geometry.
The readout question asks which parts of that geometry matter to a receiver.
Readout and Trajectory
The Cognitive Art article What Is a Trajectory? owns the path of state through time.
A readout need not depend on one static population snapshot.
It can depend on:
- where the trajectory is now,
- which direction it is moving,
- how fast it is moving,
- whether it crossed a boundary,
- the history of recent states.
A dynamic readout therefore consumes a trajectory rather than a point.
Readout and Rate Code
The Cognitive Art article What Is a Rate Code? owns information represented by spike count or firing rate over a chosen window.
A downstream readout may integrate firing rate.
For example, synaptic currents summed over tens or hundreds of milliseconds can approximate a rate-sensitive readout.
But rate information is useful only if the downstream circuit integrates over an appropriate window before behaviour occurs.
Readout and Temporal Code
The Cognitive Art article What Is a Temporal Code? owns information in timing, latency, order, interval, burst or phase beyond count alone.
A temporal code becomes functionally meaningful only if a downstream receiver can distinguish those timing differences.
A 2-ms distinction found by an external decoder is not biologically useful if the receiving circuit averages everything over 100 ms.
Conversely, coincidence-sensitive circuitry can make very small timing differences highly potent.
Temporal coding and readout therefore meet at the receiver’s integration properties.
The Behavioural Deadline
A readout that arrives after the behaviour cannot have caused that behaviour.
This sounds obvious.
It is easy to violate analytically.
Suppose a decoder predicts a choice from neural activity averaged over 500 ms.
The animal committed to the action after 180 ms.
The later activity may still correlate with the choice.
It cannot be the operative readout that produced the decision at 180 ms.
Time is therefore a hard causal boundary.
Readout and Threshold
The Cognitive Art article What Is Threshold? owns the crossing criterion.
A decision readout may integrate a population signal until a threshold is crossed.
Readout owns the mapping from population activity into the decision variable.
Threshold owns the criterion applied to that variable.
This separation prevents a common conceptual collapse:
the decoder, accumulator and decision threshold are not automatically the same mechanism.
Readout and Gain
The Cognitive Art article What Is Gain? owns sensitivity of output to changes in input.
A readout can change gain without changing which dimensions it reads.
Same weight direction.
Steeper output response.
Alternatively, attention or learning can rotate the effective readout weights so a different dimension becomes potent.
Sensitivity and selection are different operations.
Readout Can Change With Context
The same stimulus appears under two task rules.
Rule A asks:
is it red?
Rule B asks:
is it moving?
The sensory population may contain both colour and motion information.
The task requires different readouts.
Flexible cognition can therefore be partly a problem of changing which available dimension is routed into the current decision.
Readout Learning
Suppose the relevant population representation already exists.
Learning can still fail if the downstream circuit has not learned how to read it.
This creates two distinct learning problems:
- change the representation,
- change the readout.
A behaviour can improve because upstream representations become cleaner.
Or because downstream weights become better aligned with an already useful representation.
These mechanisms can produce similar behavioural improvement and require different experiments to distinguish.
The Credit-Assignment Problem
If an action is wrong, which readout weight should change?
The nervous system must somehow assign error to the connections that contributed to the output.
This is a biological version of credit assignment.
Plasticity can be shaped by:
- local pre/post-synaptic activity,
- neuromodulatory signals,
- reward prediction errors,
- recurrent dynamics,
- eligibility traces.
Neural Readout does not claim one universal learning rule.
It marks where the problem lives:
the weights connecting representation to consequence must be established, maintained and sometimes revised.
The Same Decoder Can Be Wrong for the Biology
Researchers often compare decoder performance.
Linear decoder A reaches 80%.
Deep neural decoder B reaches 95%.
Should we conclude the brain uses decoder B?
No.
The more complex decoder tells us that more information is mathematically extractable.
The biological readout question still requires anatomy, timing, plausible computation and perturbation.
Accuracy is not mechanism.
Anatomy Constrains Readout
A downstream area cannot read a neuron that does not project to it, unless the influence travels through some intermediate route.
This means anatomy supplies a hard feasibility constraint.
Before proposing a readout, ask:
- which neurons project to the receiver?
- what synaptic signs and strengths exist?
- what laminae or cell types receive the input?
- what delays separate source and target?
- what recurrent circuitry transforms the input afterward?
The 2025 parietal-output study is especially valuable because it measures population coding in identified output pathways rather than treating every recorded neuron as equally available to the same downstream reader.
Timing Constrains Readout
The source population changes.
The downstream target receives the signal later.
Synaptic integration takes time.
Recurrent processing adds more delay.
A proposed readout must fit inside the available behavioural window.
This connects directly to Rate Code and Temporal Code:
a code is useful only at a temporal precision and latency the receiver can actually exploit.
Noise Correlations Matter to Readout
Two neurons fluctuate together from trial to trial.
Is that harmful?
It depends on the readout direction.
If correlated noise lies along the same dimension that separates two stimuli, discrimination can worsen.
If it lies mostly in an irrelevant null direction, the chosen output can remain robust.
The Nature Reviews Neuroscience review The Structures and Functions of Correlations in Neural Population Codes emphasises that the effect of correlations depends on their structure relative to the signal and readout geometry.
Readout Can Ignore Large Neural Changes
Population activity changes dramatically.
Behaviour barely changes.
That seems paradoxical if one thinks every neural difference must matter equally.
Readout geometry resolves the paradox.
The change can occur largely in output-null dimensions.
Large neural distance does not guarantee large behavioural consequence.
Direction matters.
Small Neural Changes Can Matter Enormously
The opposite is also possible.
A tiny population change aligned precisely with an output-potent direction can shift a decision.
This gives an important systems lesson:
effect size in neural state space depends on alignment with the receiver, not only on Euclidean magnitude.
Readout and Causality
The Cognitive Art article What Is Causality? owns the general question of intervention and counterfactual dependence.
Neural Readout applies that standard to representation.
If population dimension X is claimed to drive output Y, a strong test changes X while controlling competing dimensions and asks whether Y changes as predicted.
Ideal perturbations would distinguish:
- output-potent directions,
- output-null directions,
- rate changes,
- timing changes,
- upstream representation changes.
Real experiments are rarely perfect.
The principle still disciplines the claim.
Perturb the Receiver, Not Only the Sender
Stimulate the upstream population.
Behaviour changes.
That shows causal influence.
It does not fully identify the readout mechanism.
A stronger programme also tests the receiver:
- silence the downstream target,
- alter its synaptic integration,
- change the relevant projection pathway,
- test whether the predicted output dimension disappears.
Readout is a relation between sender and receiver.
Studying only one end leaves the relation underdetermined.
Readout and Redundancy
Several neurons can contribute similar information to the same readout.
This can make the output robust to loss of individual neurons.
But redundancy at the encoding level does not guarantee redundancy at the projection level.
If only one subset projects to the relevant receiver, that subset has privileged functional relevance for that output.
Readout and Sparse Coding
Sparse coding remains nested inside Population Code rather than becoming another URL in this corridor.
A sparse representation may simplify a readout because relatively few neurons are strongly active.
A dense distributed representation can also be easy to read if the relevant conditions are linearly separable.
Sparsity is therefore not the same as readout difficulty.
Readout and Linear Separability
Suppose two behavioural classes occupy different regions of population space.
If one hyperplane separates them, a linear readout can classify the states.
This is linear separability.
It is best treated as a property or test of representational geometry rather than a standalone Cognitive Art owner.
Mixed Selectivity can create the geometry.
Neural Readout owns whether a downstream weighted sum can exploit it.
Readout and Decision
The existing How Decision Making Works page remains the broad decision owner.
Neural Readout does not replace it.
Instead it answers a narrower mechanistic question:
how can one available neural representation be converted into the variable that a decision mechanism receives?
The readout may feed evidence accumulation.
It may feed an action selector.
It may directly drive motor output.
The receiver defines the job.
Readout and Action
A motor command is one especially concrete readout.
Population activity in cortex is transformed into descending signals that eventually alter muscle activation.
But the route contains many stages.
- cortical population,
- descending pathway,
- brainstem or spinal circuit,
- motor neuron pool,
- muscle.
Each stage can have its own readout geometry.
There is no obligation for one single transformation to explain the entire route.
Readout and Communication Between Brain Areas
Readout is not limited to muscles.
One brain area reads another.
A projection can transmit selected population dimensions while leaving others functionally silent for that target.
This makes inter-area communication a routing problem:
which dimensions are made potent for which receiver at which time?
Readout Can Be Gated
A pathway exists anatomically.
That does not mean its influence is constant.
Inhibition, neuromodulation, oscillatory phase and recurrent state can change effective communication.
The same sender activity can therefore have different downstream impact at different moments.
Readout weights can be structurally fixed yet functionally gated.
Readout and Phase
The Cognitive Art article What Is Phase? owns position within a genuine oscillatory cycle.
If downstream excitability varies across phase, the same presynaptic spikes can have different effects depending on when they arrive.
Phase can therefore gate readout.
But this claim requires a real oscillatory component and a demonstrated phase-dependent receiver response.
No oscillator, no phase readout.
Readout and Synchronisation
The Cognitive Art article What Is Synchronisation? owns stable timing relationships.
Synchronous spikes can produce larger coincident downstream input than the same spikes spread across time.
This can make synchrony output-potent even when mean firing rate is unchanged.
Again, one must test the actual receiver.
Synchrony is not automatically functional simply because it is measurable.
Readout and Robustness
The Cognitive Art article What Is Robustness? owns performance under perturbation and variation.
A robust readout should tolerate irrelevant population changes while remaining sensitive to task-relevant ones.
In geometric terms:
- irrelevant variability should fall mainly into null dimensions,
- decision-relevant signal should remain potent.
This is a powerful design principle in both nervous systems and engineered classifiers.
Readout and Invariance
The Cognitive Art article What Is Invariance? owns what remains stable despite transformation.
A good object-recognition readout may need to ignore:
- position,
- lighting,
- size,
- viewpoint.
Those nuisance changes can move population activity.
The desired readout should remain stable.
This means nuisance variation should ideally align with dimensions the readout discounts.
Readout and Affordance
The Cognitive Art article What Is Affordance? owns action possibilities available to an agent in context.
A visual representation can contain enormous detail.
The action system may read only what matters for reaching, grasping or avoiding.
Readout is therefore one place where representation becomes action-relative.
The world may be richly represented.
The receiver selects what matters now.
Readout in Brain–Computer Interfaces
Brain–computer interfaces make readout unusually concrete.
An engineered decoder receives neural activity and generates:
- cursor velocity,
- robotic-arm motion,
- speech synthesis,
- selection commands.
This is an external readout.
Its success proves that a chosen neural representation can support a useful downstream mapping.
It does not prove the native nervous system uses the identical weights.
Readout in Artificial Neural Networks
Artificial neural networks often end with a readout layer.
A hidden representation is mapped into:
- class logits,
- a regression output,
- next-token probabilities,
- control actions.
This structural analogy is useful.
It should not be mistaken for proof that biological brains implement the same training rule, layer architecture or objective function.
Representation Versus Functionality
The 2026 Nature Reviews Neuroscience framework is useful because it prevents “representation” from becoming one vague all-purpose word.
A neural response can be sensitive to feature X.
It can contain decodable information about X.
It can even be invariant to nuisance features.
Still, the strongest functional claim asks whether downstream systems use that response in relation to X.
Neural Readout is therefore the bridge from:
information exists
to
information matters to the next computation.
The Readout Sufficiency Test
Suppose population dimension d is proposed as the readout for behaviour B.
A useful test sequence is:
- Show that d contains information about the task variable.
- Show that d is available before the behavioural deadline.
- Identify a plausible downstream receiver.
- Show that the receiver’s activity depends on d.
- Show that variation orthogonal to d matters less for the selected output.
- Perturb d while controlling competing dimensions where feasible.
- Show the predicted downstream and behavioural consequence.
No single experiment needs to do everything.
But stronger functional claims require more of this chain.
The Alternative-Readout Test
Many neural patterns correlate with behaviour.
Which one is actually read?
Compare candidate readouts.
- rate-only,
- timing-sensitive,
- one projection-defined subpopulation,
- another projection-defined subpopulation,
- static linear,
- dynamic nonlinear.
Then test which candidate best predicts downstream neural activity, causal perturbation results and behaviour.
Decoder accuracy alone should not choose the winner.
The Null-Space Test
If a readout model is correct, it predicts directions of neural change that should leave the output unchanged.
That prediction is valuable.
Instead of testing only what changes behaviour, test what should not.
A strong readout theory predicts both potent change and null change.
This is a direct application of Cognitive Art’s Counterexample and Robustness logic.
The Receiver Test
Ask five receiver questions:
- Who receives the signal?
- Which source neurons actually project there?
- What temporal features can the receiver integrate?
- What nonlinearities or gates transform the input?
- What output from that receiver changes the next computation?
Without a receiver, “readout” remains metaphorical.
The Neural Readout Audit
- Readout of which source population?
- Into which downstream receiver?
- What exact output variable is produced?
- Which source neurons or dimensions are anatomically available?
- What weights or transformation define the mapping?
- Is the mapping linear, nonlinear or dynamic?
- Which dimensions are output-potent?
- Which dimensions are output-null?
- Does the signal arrive before the behavioural deadline?
- Does perturbation of the proposed readout change downstream activity or behaviour?
- Could a different readout explain the same decoder result?
- Is the claim information availability or demonstrated biological functionality?
Failure 1: Decoder Accuracy Becomes Mechanism
An external classifier predicts behaviour accurately, so its feature weights are treated as the brain’s readout.
Repair: identify anatomical receivers, feasible timing and causal pathway evidence.
Failure 2: Every Recorded Neuron Is Assumed Available
A decoder pools neurons regardless of projection target.
Repair: distinguish recorded population from receiver-accessible population.
Failure 3: Null Means Inactive
Output-null activity is described as useless or silent.
Repair: remember that null is defined only relative to one readout and can support internal computation.
Failure 4: Potent Means Globally Causal
A dimension affects one output and is treated as universally functionally important.
Repair: state the receiver and output explicitly.
Failure 5: Late Activity Explains Early Behaviour
A decoder uses neural activity occurring after commitment.
Repair: constrain the analysis window by behavioural latency.
Failure 6: Linear Readout Is Assumed Because It Is Convenient
A linear classifier is easy to fit and becomes the presumed biological transform.
Repair: compare plausible nonlinear and dynamic alternatives.
Failure 7: Correlation Becomes Functionality
A population tracks feature X and is therefore said to functionally represent X.
Repair: add downstream-use evidence and intervention where appropriate.
Failure 8: One Readout Is Assumed for Every Task
The same fixed weights are used to explain behaviour across changing contexts.
Repair: test task-dependent routing, gating and weight changes.
Repair Path
- Define the upstream population and represented variable.
- Identify anatomically plausible downstream receivers.
- Measure source–target timing and behavioural deadline.
- Fit the simplest plausible readout model.
- Separate output-potent from output-null population dimensions.
- Compare alternative receiver-specific models.
- Test generalisation on held-out trials and contexts.
- Perturb source dimensions and receiver pathways where feasible.
- Distinguish information availability, functional use and mechanism in the conclusion.
- Revise the readout if a simpler or better-supported receiver model explains the data.
Neural Readout in Education: Analogy Only
A student produces a rich body of evidence:
- written answers,
- oral explanations,
- error patterns,
- speed,
- transfer performance.
An examination reads only part of that state.
A multiple-choice paper may be sensitive to final selections and blind to explanation quality.
An oral examination uses a different readout.
Same learner.
Different receiver.
This is not neuroscience.
It is a faithful systems analogy:
what a system contains and what another system is able to read from it are different questions.
Neural Readout in Organisations
An organisation contains thousands of measurements.
Leadership watches five KPIs.
Those KPIs become the practical readout of the organisation.
If the readout ignores queue length, queue length can deteriorate without changing the decision process until failure becomes visible elsewhere.
The state existed.
The receiver did not read it.
The lesson again transfers:
a system is partly governed by which of its many internal dimensions its decision-makers make output-potent.
Research Notes and Further Reading
For the current conceptual distinction between neural sensitivity, specificity, invariance and downstream functionality, see Pohl and colleagues, Clarifying the Conceptual Dimensions of Representation in Neuroscience (Nature Reviews Neuroscience, 2026).
For output-potent and output-null dimensions and their computational roles, see Churchland and Shenoy, Preparatory Activity and the Expansive Null-Space (Nature Reviews Neuroscience, 2024).
For the classic motor-cortex demonstration of preparatory activity occupying output-null dimensions, see Kaufman and colleagues, Cortical Activity in the Null Space: Permitting Preparation Without Movement (Nature Neuroscience, 2014).
For target-specific population coding in identified cortical output pathways, see Specialized Structure of Neural Population Codes in Parietal Cortex Outputs (Nature Neuroscience, 2025).
For the foundational population-code framework and linear extraction of functions from distributed tuning, see Pouget, Dayan and Zemel, Information Processing With Population Codes (Nature Reviews Neuroscience, 2000).
For how correlation geometry affects information available to population readouts, see The Structures and Functions of Correlations in Neural Population Codes (Nature Reviews Neuroscience, 2022).
World Return
A neural-readout model earns trust when it predicts not only which neural patterns correlate with an output but which receiver-accessible population dimensions actually change downstream activity and behaviour under intervention.
A strong model should also predict what can change without consequence.
That is the importance of the null space.
If an external decoder can extract a variable but the proposed receiver lacks the anatomy, timing or sensitivity to use it, the functional claim must weaken.
Final Thought: Information Becomes Function Only When Something Reads It
A population can contain many possible answers.
The receiver chooses which answer can matter.
Not consciously.
Through anatomy, weights, timing, inhibition, dynamics and state.
The code is possibility.
The readout is consequence.
A neural readout is the moment a pattern stops being merely present in the brain and becomes part of what the brain does next.