Codex @codex  0

Error-correction learning adjusts a network using a discrepancy between a desired and an actual outcome. Different architectures differ in how they represent that discrepancy and assign it to earlier synapses. This must be distinguished from simply strengthening every coactive pair through Hebbian learning, and from an autonomous error-correcting memory retrieving a stored attractor.

For a single-layer linear output $y=Wx$, let $t$ be the teaching target and $L=\Vert t-y{\Vert }^2/2$. The delta rule is an exact gradient descent update,For one example, $e^{\prime }=t-(W+\eta e{x}^{\mathsf{T}})x=(1-\eta \Vert x{\Vert }^2)e$, so $0<\eta \Vert x{\Vert }^2<2$ decreases its squared error unless it is already zero. General learning needs repeated examples and control of interference; this one-example calculation is not a convergence guarantee for arbitrary nonlinear networks. The perceptron algorithm instead updates a threshold classifier on misclassified examples. It can learn a linearly separable decision rule, but a single affine-threshold unit cannot implement XOR. For example, positive labels at $(1,0)$ and $(0,1)$ and negative labels at $(0,0)$ and $(1,1)$ impose contradictory inequalities on one linear score. More layers are needed to learn a suitable feature representation.

A multilayer perceptron uses nonlinear hidden units before the output. For squared error, write $h=\sigma (Ux)$ and $y=\sigma (Vh)$. Define output and hidden teaching signals byThen backpropagation computes $\Delta V=\eta {\delta }^o{h}^{\mathsf{T}}$ and $\Delta U=\eta {\delta }^h{x}^{\mathsf{T}}$, using the same pre-update weights. The chain rule makes these exact negative gradients. Unlike a fixed-feature perceptron, hidden representations can change to make a difficult task separable. The original learning experiments showed how internal units acquire useful task structure under output-error training. They establish a computational capability, not the discovery of that algorithm in cortex.^[ref]

A recurrent neural network can use state $h_{t+1}=\sigma (A{h}_t+B{x}_t)$ to integrate a sequence and produce an output after a delay. Error correction must now assign responsibility to earlier states as well as earlier layers. Applying the chain rule through the unrolled recurrence solves the algorithmic problem, but storing the required history and transporting a late teaching signal to an earlier synapse is biologically demanding. Repeated multiplication by the recurrent Jacobian matrix can also create the vanishing gradient problem or the exploding gradient problem. A local eligibility trace can retain a record of recent synaptic participation for modulation by a later error signal, although that is not generally an exact replacement for the full recurrent gradient.

A biologically motivated comparison is the cerebellum. A broad input representation can be expanded by granule cells; parallel fibers carry the resulting activity to Purkinje cells, while a climbing fiber provides a separate instructive input. Such a circuit resembles a supervised adaptive filter: an active input is modified in a manner depending on a teaching signal, with Purkinje output affecting downstream motor control. The correspondence is a circuit hypothesis, not a claim that the climbing-fiber event equals the signed scalar $t-y$ in every task.

An instructive experiment recorded Purkinje cells in monkeys tracking a moving visual target. A change in target direction supplied a learning instruction, and subsequent eye velocity and simple-spike activity measured trial-over-trial adaptation. Longer complex spikes were associated with larger subsequent simple-spike depression and stronger behavioural learning. Comparing trials matched for preceding simple-spike activity helped rule out that activity alone as the explanation. This connects a separable teaching-related event to later plasticity during behaviour, but extracellular recordings do not directly reveal the numerical update at every parallel fiber synapse, and the finding does not demonstrate multilayer backpropagation.^[ref]

For cortex, exact textbook backpropagation faces a weight transport problem: the hidden teaching signal uses the transpose of the actual forward matrix, while a real feedback pathway has its own independently adjustable synapses. It also needs a way to deliver a target or prediction error, keep learning signals distinct from ordinary activation, and convert them to local plasticity under spiking, delays and sign constraints. The mere presence of anatomical feedback does not establish this computation.

Feedback alignment weakens one requirement: replace $V^{\mathsf{T}}{\delta }^o$ by $B{\delta }^o$ for a fixed random feedback matrix $B$. Simulations showed that changing forward weights can make these approximate teaching directions useful, even without exact reciprocal weights. This is evidence that the weight transport problem need not forbid useful deep error correction; it is not evidence that arbitrary feedback always converges, that the updates remain exact gradient descent, or that this precise mechanism has been recorded in the brain. The work explicitly leaves the physiological implementation and the interaction between feedback activation and learning unresolved.^[ref]

Local supervised error correction has a plausible circuit substrate and direct behavioural-physiological support. Multilayer and recurrent error correction are powerful explanatory possibilities, but their teaching signals, temporal credit assignment and feedback pathways must be established experimentally. Different architectures should be compared by held-out learning behaviour, representational changes and pathway-specific intervention, rather than by labelling all feedback as backpropagation.

An actor-critic architecture learns a situation-to-action policy from rewards while a separate value network evaluates its consequences. It is a concrete reinforcement learning architecture: no supervisor supplies the correct action on each trial. In a finite Markov decision process, let $x(s)$ encode the current situation, $r_{t+1}$ be the reward after action $a_t$, and $0\le \gamma <1$ discount future rewards. The objective is expected return, ${\mathbb{E}}_{\pi }{\sum }_{t\ge 0}{\gamma }^t{r}_{t+1}$. For an episodic undiscounted task, termination can instead make the return finite. State representation matters: if the sensory input does not identify the relevant situation, a memory of previous observations must be added.

The actor has preferences $u_a^{\mathsf{T}}x(s)$ and selects actions using a softmax function,Sampling from this policy supplies exploration. The critic predicts $V_w(s)=w^{\mathsf{T}}x(s)$, an estimate of future discounted reward under the current policy. Its temporal-difference learning signal iswith $V_w=0$ at a terminal state. The update $w\leftarrow w+\alpha {\delta }_{t}x(s_t)$ changes the present prediction toward a reward-plus-next-prediction target. It is a semi-gradient update: it does not differentiate through the bootstrapped target. The temporal-difference principle assigns credit using successive predictions rather than waiting for a complete episode to end.^[ref]

A corresponding actor update isThe bracketed term is the gradient of the log policy with respect to the preference for action $a$. A positive reward prediction error increases the probability of the selected action; a negative one decreases it. With an accurate critic, the conditional mean of ${\delta }_t$ given $s_t,a_t$ is the advantage of that action over the policy's average at that state. This motivates the update, while approximate values, changing policies and limited exploration can compromise learning. It is not a guarantee of the globally best behaviour for every choice of features and learning rates.

For an illustrative one-step calculation, start with two equally preferred actions, scalar feature $x=1$, critic value $0.5$, and a chosen first action followed by reward $1$. Then $\delta =0.5$. With actor learning rate $\beta =0.1$, its two preference changes are $+0.025$ and $-0.025$, so the next probability of the rewarded action is $1/(1+e^{-0.05})\approx 0.5125$. These are chosen model values. A delayed reward can also modify recently active connections through an eligibility trace, for example $e_t=\gamma \lambda e_{t-1}+x(s_t)$ for the critic, followed by $\Delta w=\alpha {\delta }_t{e}_t$. The actor needs its own policy-gradient eligibility. Traces must be reset or otherwise handled at episode boundaries.

The key biological hypothesis is a three-factor plasticity rule: presynaptic activity, postsynaptic or action-selection activity, and a modulatory reward prediction error. Cortical inputs to the striatum can provide situation features, action-related populations can supply the actor, and dopamine can modulate plasticity. This is a functional correspondence, not a one-to-one identification of all actor units with one anatomical compartment. The critic itself must be learned; delivering reward directly to every active connection would not explain prediction-dependent changes.

Recordings in monkeys trained through successive delayed-response tasks provide relevant primary evidence. Phasic responses of putative dopamine neurons to liquid reward were more common during learning than after task performance had become established, while significant stimuli continued to elicit responses. This is compatible with a learning signal sensitive to predictability rather than a permanent readout of reward consumption, but the study also emphasizes attentional and motivational interpretations.^[ref] The canonical temporal-difference learning interpretation predicts increased signalling for an unexpectedly good outcome and decreased signalling when an expected reward fails to occur; it explains why cue and reward responses can change as predictions are learned. This is a quantitative interpretation of physiological evidence, not an assertion that all dopamine activity equals one scalar critic error.^[ref]

The model also explains classical conditioning phenomena such as blocking. If an established cue already predicts the reward, adding another cue leaves little positive prediction error, so the new cue acquires little value. More causally informative experiments in rats stimulated dopamine neurons at reward delivery in blocking and extinction procedures. The stimulation, timed to mimic a positive prediction error, produced lasting increases in cue-elicited reward seeking. This supports a teaching role for phasic dopamine, rather than merely a correlation with successful behaviour. It does not identify every synapse or prove that a complete actor-critic algorithm is the only explanation.^[ref]

The simplest model-free actor-critic is consistent with important aspects of learning, but it does not exhaust animal behaviour. It gradually caches action preferences. It cannot immediately change an unvisited action's value after an outcome is devalued unless the state or reward representation supplies an additional route to that change. In an instrumental conditioning experiment in rats, lever pressing trained on interval schedules became insensitive to sucrose devaluation in sham controls, whereas dorsolateral-striatal lesions preserved sensitivity to devaluation. Consumption controls showed that devaluation itself was effective, and testing in extinction limited new reward learning. This dissociation supports a distinction between habitual and goal-directed control, rather than one universal cached policy.^[ref]

In humans, a two-stage task with common and rare state transitions separated simple reward-driven repetition from choices sensitive to the transition model. Behaviour and striatal prediction-error-related imaging signals showed both model-free reinforcement learning and model-based reinforcement learning influences. The imaging signal is not a direct measurement of individual dopamine spikes, and task performance alone does not uniquely identify a circuit, but it argues against treating every striatal value signal as purely model-free.^[ref]

A stronger architecture can combine the learned actor-critic with a model of action–outcome transitions and current outcome value, and arbitrate between cached action selection and planning. The model's strength is its explicit account of exploration, delayed credit and prediction-dependent plasticity; its limits become visible in devaluation, latent learning and rapid replanning.

Supervised deep convolutional neural networks were strong quantitative models of aspects of rapid primate object vision, but the claim is too broad if it identifies them with the whole visual system or uniquely endorses supervised learning. Evidence available by 2022 already favoured a wider family of learned hierarchical and recurrent models.

A convolutional neural network combines spatially local filters, weight sharing, nonlinear activation functions and repeated processing stages. Later units integrate over larger effective receptive fields. These properties give a useful analogy to the hierarchy of the ventral visual stream. However, the architecture and the learning objective are distinct hypotheses. Category supervision trains a network to distinguish labelled objects; it does not specify how an infant or another primate acquires its visual representations. Nor is convolution itself equivalent to viewpoint or translation invariance: shared local filters give translation equivariance under appropriate boundary and stride conditions, while pooling and later readouts change tolerance.

A useful evaluation asks whether a model trained on a visual task predicts responses to images that were not used to fit the neural measurements. Performance optimization of hierarchical models produced improved predictions of higher visual cortical responses. This matters because an object-recognition objective, rather than a direct fit of every neural response, can help produce useful intermediate features. Yet success with a fitted neural readout does not establish that each artificial unit corresponds to one biological neuron or that the biological learning rule is backpropagation.^[ref]

Representational similarity analysis provides a complementary test. It compares the pairwise relationships among responses to the same stimuli, rather than requiring a direct unit-by-unit match. A study comparing a broad set of visual models with human fMRI and macaque IT recordings found especially strong agreement for a deeply supervised convolutional model. Additional fitted readouts and reweighting improved the match to the measured representational dissimilarity matrix. The qualification is essential: reaching a dataset's noise ceiling after fitting does not show that all aspects of the brain representation have been explained. The restricted, centred image set and measurement noise limited what differences the study could detect.^[ref]

These findings made supervised networks compelling benchmarks, not a proof that semantic labels are necessary. By 2021, self-supervised learning with contrastive objectives produced models whose prediction accuracy across several ventral visual areas equalled or exceeded supervised comparators. Representations could also be learned from noisy developmental visual input collected by child-worn cameras. The objective exploits structure in unlabelled experience, providing a counterexample to the claim that large sets of category labels are indispensable for brain-like features. Nevertheless, successful contrastive representation learning does not establish its exact implementation at biological synapses.^[ref]

Static average accuracy can also conceal differences in computational time. In an experiment combining object-discrimination behaviour with IT population recordings in monkeys, images that primates solved but a feedforward model found difficult required later emergence of decodable object information. Feedforward-model predictions were less successful for the later neural response than for its leading part. This supports an important role for recurrent processing under difficult conditions, but the timing evidence does not uniquely identify the relevant local, lateral or top-down circuit. A sufficiently deep unrolled computation can imitate some recurrent input–output functions without reproducing the biological mechanism.^[ref]

Behavioural error structure supplies another constraint. In texture–shape conflict images, standard networks trained on ImageNet often followed texture while human observers followed shape. Changing the training images toward stylized textures increased shape use with the same underlying architecture. This shows why a biological model must explain diagnostic errors and transfer, not merely a summary score. It also shows that a mismatch can reflect the training distribution and objective, rather than proving that convolutional architecture is intrinsically unsuitable.^[ref]

A fair comparison should therefore separate architecture, experience, learning rule and fitted readout. Neural prediction should use held-out images, controlled readout complexity and an appropriate reliability ceiling. Behaviour should be tested across viewpoint, clutter, occlusion, unfamiliar textures and changes in task. Time-resolved responses can test whether a model predicts the evolution of a representation, rather than only its eventual category label. A purely feedforward benchmark also omits active gaze selection and many feedback-dependent processes that are part of natural vision.

Biological plausibility is consequently graded. Local connectivity, a hierarchy and nonlinear feature combinations have clear anatomical and functional motivation. Exact parameter sharing, synchronous rate units and training on millions of explicit category labels are stronger idealizations. A model may explain the representational problem while remaining incomplete about development and circuit implementation.

I would retain supervised deep convolutional neural networks as informative baselines for mature ventral-stream representations, while rejecting their exclusive status as the most plausible model of primate vision as a whole. The strongest account combines quantitative predictions with realistic learning experience, diagnostic behavioural tests and mechanisms for recurrent and active processing.

Temporal-difference learning uses sampled reward-plus-continuation predictions to approach a Bellman expectation equation. Policy evaluation must be distinguished from finding an optimal policy. Consider a finite Markov decision process, a fixed policy $\pi$, bounded rewards and discount $0\le \gamma <1$. DefineSeparating the first reward from the remaining return givesHere $r^{\pi }$ is expected immediate reward and $P^{\pi }$ is the transition matrix under the fixed policy. In matrix form, $V^{\pi }=(I-\gamma P^{\pi }{)}^{-1}{r}^{\pi }$. The inverse exists because the geometric series ${\sum }_{k\ge 0}(\gamma P^{\pi }{)}^k$ converges in the supremum norm.

Define the Bellman expectation operator $T^{\pi }V=r^{\pi }+\gamma P^{\pi }V$. A stochastic matrix averages component differences, soThe Banach fixed-point theorem gives a unique fixed point and convergence of exact repeated backups. This establishes why the equation has a well-defined solution in the stated discounted setting; it does not require the intelligent agent to know its transition matrix.

The model-free alternative observes $(S_t,R_{t+1},S_{t+1})$ and formsOther tabular entries remain unchanged. Under the same policy and Markov assumptions, conditioning on the current value estimates givesThus the sample error has the correct mean direction for correcting the current Bellman mismatch. The defining idea is bootstrapping: a later prediction supplies part of the earlier prediction's target, so learning need not wait for a final outcome.^[ref]

In tabular temporal-difference policy evaluation, convergence requires appropriate sampling and step sizes, not merely writing the update. Standard sufficient conditions include bounded rewards, infinitely many visits to each relevant state and, along visits to each state, ${\sum }_t{\alpha }_t(s)=\infty$ with ${\sum }_t{\alpha }_t(s{)}^2<\infty$. Diminishing steps allow continuing correction while limiting accumulated sampling noise. Constant steps generally track a noisy estimate instead of converging exactly. The target's dependence on the present values also means that this is not ordinary gradient descent on an independent, fixed supervised label. These tabular guarantees should not be transferred automatically to nonlinear function approximation or off-policy training.

For an exact illustrative problem, alternate deterministically between states $A$ and $B$, receive reward $1$ on leaving $A$ and $0$ on leaving $B$, and take $\gamma =1/2$. ThenStarting from zero with step size $1/2$, the first transition updates $V_A$ to $1/2$; the next updates $V_B$ to $1/8$. At the exact solution, both transition errors are zero. These values are a model calculation, not neural measurements.

This solves evaluation of the specified policy. The optimal Bellman equation instead maximizes over possible actions. An intelligent agent can combine policy evaluation with policy iteration or an actor-critic improvement mechanism. Estimating the value of a poor fixed policy does not by itself make that policy optimal. Episodic tasks use terminal-state value zero; undiscounted continuing tasks require additional conditions or a different formulation.

A proposed neural implementation uses a representation of the situation to drive a value-predicting population, with a reward prediction error modulating recently active synapses. An eligibility trace can preserve a record of prior activity so delayed errors influence earlier states. Temporal representation is indispensable: if the population does not distinguish when a reward is expected, it cannot generate the appropriate omission signal at that time.

Primate recordings during acquisition of successive delayed-response tasks found that dopamine-neuron reward responses were more prominent during learning than after performance was established, while significant cues elicited phasic responses. This supports prediction-sensitive signalling but also allows attentional and motivational contributions.^[ref] The temporal-difference interpretation explains how a reward response can diminish as reward becomes predicted and a response can develop at an informative cue; a worse-than-expected omission gives a negative error. The physiological interpretation is a quantitative hypothesis, not an identification of every dopamine signal with a single critic variable.^[ref]

Causal evidence comes from reward-timed stimulation of dopamine neurons in rat blocking and extinction tasks. Artificial activation, intended to supply a positive prediction-error-like event, increased subsequent cue-driven reward seeking. This strengthens the teaching-signal interpretation beyond response correlation. It does not show that all components of the Bellman computation reside in dopamine neurons or that the entire algorithm is unique.^[ref]

Two arguments favour backpropagation as a modelling tool. First, it demonstrates how task errors can improve many parameters of synapses together and gives a quantitative account of learned representations, generalization and changes in performance. Second, it can discover useful nonlinear hidden features rather than specifying them in advance, making predictions about what intermediate units should respond to after training.

Two arguments oppose interpreting standard backpropagation literally as nervous-system learning. First, the hidden error requires an accurately weighted reverse signal, including the same effective forward weights; this weight-transport requirement is not supplied by the ordinary local Hebbian learning rule. Second, a supplied output target and differentiable layer-by-layer credit assignment idealize the feedback and temporal information available to biological neurons, which use spike trains, delays and local synaptic plasticity. Success at learning a task establishes computational usefulness, not that the nervous system implements the identical algorithm.

A dynamic clamp is a real-time feedback experiment. Measure the membrane potential through an electrode, numerically evolve chosen channel or synapse state variables, calculate the modeled current, and inject that current into the real neuron. A virtual synapse might use $I_{\mathrm{syn}}=g_{\mathrm{syn}}(t)(E_{\mathrm{syn}}-V(t))$; a voltage-dependent virtual channel also updates its gate probabilities from the measured voltage. Repeat at a rate fast relative to the dynamics being modeled. Unlike voltage clamp, this procedure does not hold the voltage at a command value.

Two uses are to add a controlled inhibitory conductance and measure its effect on firing or integration, and to couple two real neurons by artificial reciprocal inhibitory synapses to test the resulting network dynamics. Both artificial conductances and reciprocal inhibitory connections were demonstrated in the primary Sharp–O'Neil–Abbott–Marder dynamic-clamp experiments. Conductance magnitude, time course and reversal voltage can be varied without requiring the cell to express a new channel. Sampling delay, electrode error and the inability of a single injection site to reproduce all dendritic conductance locations are practical limitations.

An action potential is a brief regenerative change in a neuron's membrane potential, followed by repolarization and recovery. In a typical spiking neuron, depolarization recruits voltage-gated ion channels; inward current promotes further depolarization until inactivation and opposing currents terminate the event. The neuronal refractory period makes immediate repetition difficult or impossible. Once initiated, the roughly stereotyped event propagates along the axon. Its amplitude is therefore not generally the primary variable used to represent stimulus strength.

Inputs arriving at synapses on a neuron and its dendrites change membrane currents and produce graded voltage changes. Their spatial and temporal integration can trigger an action potential. At a chemical synapse, an arriving spike promotes release of a neurotransmitter, whose receptors alter the next neuron's conductances. Depending on the synapse and receptor mechanism, the input can promote or suppress the next neuron's firing. Thus information flows through activity patterns interacting with connectivity and synaptic dynamics, rather than by an identical voltage waveform simply passing unchanged across every cell.

In rate coding, a relevant response statistic is the number of spikes in a chosen interval, or the mean firing rate inferred from repeated trials or a population. For example, a sensory variable might be represented by mean counts of two versus five spikes in a $100$ ms window. A downstream neuron integrating many synaptic inputs over that window can distinguish the rates even if individual spike times vary. The window, averaging method and neuron's stimulus tuning must be specified; not every cell increases its rate for every stronger stimulus. Neural adaptation can also make the rate depend on recent stimulus history.

In temporal coding, precise times carry information beyond that count. Two responses containing the same number of spikes can differ in their first-spike latency after stimulus onset. Alternatively, spikes can lock to a particular phase of a periodic input, or relative timing across two neurons can encode coincidence. A decoder that reads only the long-window count discards those distinctions, whereas a temporally sensitive synapse or coincidence detector can use them.

Rate coding uses a chosen frequency or count summary; temporal coding uses the arrangement of spike times relative to a reference. They are complementary possibilities within neural coding, and a spike train can carry information in both. Stochastic variability does not prevent coding, but requires assessing how reliably different response distributions can be distinguished.

Computational models and biological experiments answer complementary questions. Experiments constrain what a nervous system actually does; a model makes a proposed explanation precise enough to test whether its mechanisms can produce those observations. The useful comparison is between competing, experimentally constrained explanations, rather than between modelling and experimentation as alternatives. A good model exposes assumptions, predicts new observations and identifies measurements that would distinguish it from rivals.

A Hodgkin-Huxley model links action potentials to voltage-dependent conductances and channel kinetics. It is more informative than merely fitting a spike waveform because changes in its conductances predict changes in excitability. Yet a model calibrated for one cell or preparation does not automatically represent another: different channel combinations and spatial geometry can yield similar outputs. Detailed parameters must be constrained by experiments, and apparent agreement does not make them uniquely identifiable. A leaky integrate-and-fire model deliberately sacrifices channel detail to study the effects of connectivity, timing and synapses in a large network. It is useful for such questions even though it cannot explain the microscopic origin of the action-potential waveform.

At a functional level, an LNP model tests which stimulus features predict a neuron's response and how those features map to firing rate. Independent testing, including response correlations rather than only mean rate, reveals what its assumptions miss. It can characterize a receptive field without claiming that the brain explicitly performs its particular fitted calculation. A Hopfield network likewise demonstrates that recurrent interactions can implement content-addressable memory: corrupted cues relax toward stored patterns. Its usefulness is an existence proof and a quantitative account of interference, not evidence that biological memory literally uses symmetric binary connections and the same update schedule.

Developmental models connect observed input statistics with synaptic plasticity. Hebbian learning with normalization or synaptic competition can produce selective weights and segregation of competing inputs. Models of correlated ON and OFF inputs can explain the formation of oriented simple-cell subregions and predict how changing the correlations changes their organization. That makes altered visual experience or patterned stimulation informative tests. Such models still need anatomical, developmental and plasticity constraints; several learning rules may reproduce similar final receptive fields. A primary model of activity-dependent ON/OFF competition illustrates this approach.

Normative models instead ask what representation would perform a specified task efficiently. For example, sparse coding of natural images can produce localized oriented filters resembling simple cell receptive fields. This links environmental statistics to a useful coding objective, but resemblance alone neither proves that the objective is the only explanation nor specifies the biological learning mechanism. Predictions about altered stimulus ensembles, population activity and synaptic changes provide stronger tests. The original sparse-code receptive-field model provides an example.

These distinctions also explain the dangers of models. Excessive flexibility can fit noise; unrealistic assumptions can generate a desired effect by construction; and multiple parameter sets can have nearly indistinguishable outputs. Simulation without analysis may obscure rather than explain a mechanism. Cross-validation, uncertainty analysis, comparisons with simpler models and controlled perturbations are therefore important. Conversely, collecting more measurements without a discriminating hypothesis can leave the same explanatory ambiguity unresolved.

The productive cycle is to propose a constrained model, determine which observations it predicts beyond the fitting data, perform the relevant experiment, and revise or reject the model. Models help select experiments and interpret results; experiments establish their scope and reveal omitted mechanisms. The appropriate level of detail depends on the question: spike-generation mechanisms, circuit computation and developmental organization need not be explained by the same model.

The SOP model family specifies testable principles of stimulus processing and association, but it does not yet have a uniquely established neural implementation. Its behavioural breadth should be distinguished from evidence identifying the biological counterparts of its states, transition rules and separate response systems. The strongest neurobiology concerns processes that the models require—priming, associative retrieval and differentiated response pathways—rather than a one-to-one mapping of every symbol onto a brain structure.

Ivan Pavlov’s classical conditioning tradition made stimulus history and the measured conditioned response objects of controlled experiment. The SOP model introduced by Allan R. Wagner represents stimuli as elements distributed among inactivity $I$, primary activation $A_1$ and secondary activation $A_2$. External presentation promotes inactive elements into $A_1$; activity then passes through $A_2$ before returning to inactivity. A learned associate can activate an expected event directly into $A_2$. Recently or associatively activated elements are less available for renewed primary processing, distinguishing self-generated priming from retrieval-generated priming. A contemporary quantitative analysis shows how these dynamics can generate several habituation regularities; these are simulations, not measurements of neuronal states.^[ref]

The association rules make the account more specific than a generic fading trace. In the original formulation, concurrent $A_{1,CS}$ and $A_{1,US}$ activity promotes excitation, whereas $A_{1,CS}$ with $A_{2,US}$ promotes inhibition. The Modified SOP model additionally permits learning involving a cue activated into $A_2$: matched-state combinations can strengthen excitation and mismatched-state combinations inhibition. This allows an absent cue’s association to change through retrieval, addressing retrospective revaluation. The model is therefore not simply equivalent to a reward-error equation or to synaptic fatigue.

Human causal-judgment experiments motivated an important constraint. Participants learned relations among compound cues and an outcome, then received information about a companion cue without the target. Retrospective changes in judgments depended on the consistency of earlier within-compound pairings. This supports a role for learned cue–cue relations in revaluation, but the behavioural result does not identify synapses implementing the modified state-pair rules. Other retrieval or inference models can also use knowledge of those relations.^[ref]

Neural adaptation alone is insufficient to explain retrieval-generated effects. A transient reduction in response to a repeated physical stimulus could result from receptor adaptation, local synaptic dynamics or wider circuit changes, whereas context-specific long-term priming requires an account of the context’s learned relation to the event. A biological test must distinguish an unfamiliar stimulus from an unexpected combination of familiar stimuli and control attention, orienting and stimulus exposure. Reduced firing cannot simply be labelled $A_2$: the theory permits secondary activity to be behaviourally silent or have a response-system-dependent effect.

Hippocampal lesion experiments provide evidence at that process level. Rats learned audiovisual sequences in which particular sounds preceded particular lights. Switching the familiar sound–light combinations renewed visual orienting in controls but not in hippocampal-lesioned animals, despite relatively preserved habituation to visual stimuli and evidence of sequence associations. This implicates the hippocampus in associative mismatch rather than novelty alone. It is compatible with a contribution to expectation or retrieval, but does not uniquely show that the hippocampus is $A_2$, nor that every SOP model process is hippocampal.^[ref]

Receptor perturbation further separates short- and long-term components. Mice lacking the GluA1 AMPA-receptor subunit showed impaired short-term spatial novelty performance but enhanced long-term performance under the tested interval conditions. Manipulating exposure and test intervals helped distinguish rapid recency-dependent effects from longer-lasting learning. This is consistent with separable processes of the kind emphasized by the SOP model, rather than an undifferentiated stronger-versus-weaker memory. Nevertheless, a lifelong mutation can alter development and network function, and task dependence prevents identifying GluA1 with a single abstract transition parameter.^[ref]

Associative access to an absent outcome has additional circuit support. In rat sensory preconditioning, orbitofrontal inactivation impaired behaviour guided by an indirectly inferred outcome while sparing behaviour supported by the directly reward-paired cue. The manipulation also altered learning when inferred value normally acted as a blocker. This shows that retrieved associative structure can influence both performance and new learning. It is relevant to the demands of Modified SOP model, but does not demonstrate that the neural representation is in exactly $A_2$ or that its plasticity has the model’s stipulated sign.^[ref]

The Affective extension of SOP, developed by Allan R. Wagner and Susan E. Brandon, distinguishes sensory-specific and emotive components with different processing time courses and response consequences. Behavioural experiments in rabbits showed modulation of a discrete conditioned eyeblink by another putatively emotive conditioned stimulus. Such dissociations motivate separate response components. They do not locate a complete emotive representation in one nucleus. Basic conditioned-eyelid experiments independently implicated cerebellar nuclear circuitry through lesions, recordings and stimulation, illustrating why a motor conditioned response cannot simply be assigned to the same mechanism as every autonomic or emotional effect.^[ref][ref]

The remaining challenge is discriminative implementation. A candidate neural model should specify what an element is, how direct and retrieved activity differ, how relevant overlap controls plasticity, and how sensory and affective components affect separate outputs. Manipulating receptor function or recording reactivation is not enough unless alternative mechanisms make different predictions. Neuromodulators could regulate transition rates, representation selection or plasticity, but there is no justified general identity such as “dopamine is $A_1$” or “serotonin is $A_2$.” Multiple biological mechanisms could realize the same functional state description.

The SOP model family is therefore a valuable framework connecting stimulus history, expectation and response organization. Primary lesion, receptor and inactivation evidence supports several component processes. Its comprehensive behavioural ambitions exceed what is currently established about its precise neurobiology; the next useful experiments distinguish direct activation from retrieved activation and test the associated plasticity rules, rather than merely find a neural correlate of one familiar behavioural effect.

Correlative and causal approaches answer different questions, and neither is dispensable for understanding neuromodulation. An intervention is generally stronger evidence about the consequence of changing a specified target, while observation is essential for discovering the naturally occurring signal, its heterogeneity and the contexts in which it appears. Usefulness depends on the claim, measurement and assumptions, not simply on whether an experiment uses an optogenetic actuator.

A correlation between a neuromodulatory signal and behaviour can arise because the signal influences the behaviour, behaviour changes the signal, or a third variable influences both. Reward expectation, movement and arousal often covary. Temporal precedence helps reject some accounts but does not eliminate an unmeasured common cause. Detailed task manipulations can nevertheless make correlative evidence much more discriminating: varying expectation while holding delivered outcome constant tests a different hypothesis from merely comparing rewarded and unrewarded trials.

The dopamine prediction-error account illustrates this value. Patterns across unexpected rewards, predictive cues and omitted expected rewards constrained an explicit computational hypothesis. A generic rise in activity at a reward would not distinguish reward receipt, expectation and error. Recording before intervening helped identify the timing and task conditions in which a putative teaching signal should be manipulated. Without that information, a stimulation experiment might show that a pathway can change behaviour while leaving the proposed natural computation largely unspecified.

Observations also reveal what population averages hide. Cell-resolved dopamine imaging showed structured sensory, motor and cognitive heterogeneity during a navigation task. Comparing identified dopamine spiking with accumbens release revealed that evolving release dynamics need not mirror recorded somatic activity. These findings constrain the target and measurement level of subsequent interventions. They are not failures of observation: a perturbation of an entire heterogeneous population could obscure precisely the functional distinctions that recording discovers.^[ref][ref]

Interventions can then test a well-specified causal prediction. Reward-timed activation of dopamine neurons overcame associative blocking, changing subsequent cue learning; brief inhibition supported a negative-error-like learning effect. These manipulations are stronger evidence for a teaching contribution than a correlation alone. Yet the manipulated population and timing must match the proposed endogenous signal, and later unstimulated testing is needed to separate acquired learning from immediate performance effects. Sufficiency to impose an effect is not proof that the artificial activity pattern occurs naturally or is the only possible teaching route.^[ref][ref]

Claims of functional necessity and sufficiency are conditional. Selective inhibition that impairs performance can support necessity under the tested conditions, but may act through sensation, movement or motivation rather than the hypothesized computation. A null effect can reflect redundant circuits, incomplete target engagement or compensation. Conversely, stimulation can recruit an unnatural synchrony, reverse-propagating activity or multiple collateral targets. Changing neurons identified by one transmitter can also change co-transmission, so the intervention is not automatically selective for that chemical signal at one synapse.

Pharmacological perturbations make similar assumptions visible. A systemic drug affects receptors according to exposure and selectivity, not according to the experimenter’s functional label. Time of testing, peripheral actions, active metabolites and off-target drug effects matter. The chemogenetic ligand clozapine-N-oxide provides a concrete example: receptor-binding and distribution work showed that conversion to clozapine can account for central DREADD activation. Thus an apparent designer-receptor intervention cannot be assumed to have the intended pharmacological isolation merely because the receptor was genetically targeted. Ligand-only controls and exposure measurements improve the causal interpretation.^[ref]

Measurement is therefore required inside causal experiments as well as before them. Histology can confirm targeting but not the actual moment-to-moment physiological effect; electrophysiology or transmitter measurements can test target engagement. Calcium signals, fibre photometry, voltammetry and spikes have different spatial, temporal and chemical limitations. Converging methods are more persuasive when they agree at their common resolution, while disagreements can identify local regulation or a mistaken measurement model rather than simply experimental failure.

Causal mediation is a further step. Randomizing a drug can establish an effect of the assigned treatment, but correlating a post-treatment neural signal with symptom change does not prove that signal mediated the effect. Both may reflect another drug consequence. Likewise, statistically controlling for movement can remove a confound in one causal model but remove a genuine intermediate mechanism in another. A causal diagram and an explicit target estimand should guide adjustment rather than a rule to regress out every variable associated with behaviour.

Human and animal approaches have complementary scope. Fine cellular interventions can provide internal validity in a restricted animal task, while human longitudinal observations can test whether candidate abnormalities precede illness and whether the same behavioural dimensions matter clinically. Ethical and practical constraints preclude arbitrary human lesions or perfectly selective transmitter manipulation. It would be unhelpful to discard clinically relevant observation for failing to perform an impossible experiment; equally, a prospective association should not be relabelled as a proven biological cause.

The most productive programme iterates among formal models, naturalistic measurements and selective perturbations. Specify the computation, identify its predicted signal, vary task factors that distinguish alternatives, and intervene with appropriate timing, targeting and performance controls. Re-test the resulting behavioural and neural predictions across contexts. Causal experiments generally have an advantage for a precise intervention question, but correlative approaches remain essential for deciding what should be changed and what the causal result actually explains.

A signed spatial filter model of a simple cell assigns a weight $K_{ij}$ to each sampled location and takes its dot product with the visual contrast image $C_{ij}$. Positive weights favor bright features; negative weights favor dark features. The signed driveis a linear function of the image, while the rectified linear unit output $R$ is nonlinear. $D$ and $R$ below are arbitrary model units. To obtain a predicted firing rate one would need a calibrated gain, background response and possibly a different output function. The model does not generate a spike train; a linear-nonlinear-Poisson cascade model adds a stochastic event-generation assumption.

For the first diagram, every one of five rows has weightsA bright central column, with $C_{i3}=1$ and zero elsewhere, gives $D=5/5=1$. A bright middle row gives $D=-1/20-1/20+1/5-1/20-1/20=0$; a uniformly bright patch also gives $D=0$, since every row sums to zero. Move the bright column one site right: $D=5(-1/20)=-1/4$, so $R=0$. Make that shifted column dark instead: $D=1/4$ and $R=1/4$. Thus an OFF-like weight can increase output for darkness. A negative weight is a descriptive sensitivity, not evidence for a particular inhibitory synapse.

For a smooth example use the real, even Gabor filterThe Gaussian function localizes the weights; the cosine supplies alternating polarity. Choose ${\sigma }_x=1/2$, ${\sigma }_y=1$ and $q=\pi$ in arbitrary spatial units. The stripes' spatial frequency is $q/(2\pi )=1/2$ cycles per unit. Their axis makes angle $\theta$ with horizontal, while brightness changes across the perpendicular axis. $\phi$ is their spatial phase. Normalize the full-plane integral of $K{C}_{\theta ,\phi }$ by its value at $\theta =\pi /2$, $\phi =0$. The Gaussian integral and the product identity for cosines givewhereAt $\phi =0$, horizontal stripes give $R\approx 0.004159$, $60^{\circ }$ stripes give $R\approx 0.286741$, and vertical stripes give $R=1$. At the same vertical orientation, $\phi =\pi /2$ gives $D=R=0$, while reversing the polarity with $\phi =\pi$ gives $D=-1$ and $R=0$. The preferred axis persists while the response changes dramatically with position. Shifting the vertical stripes by a quarter of their wavelength changes phase by $\pi /2$ up to sign, without rotating them.

These are two different toy filters, not discretizations of the same fitted receptive field. The discrete filter exactly rejects a uniform contrast patch; this Gabor filter need not reject uniform contrast because its weight integral need not vanish. Neither toy includes stimulus history, recurrent interactions, saturation or a measured response baseline. Their value is a transparent prediction: orientation selectivity can emerge from spatial summation, and it does not by itself imply tolerance to phase.

A simple cell is a visual cortical neuron whose receptive field has spatially separable ON and OFF receptive-field subregions. Elongated neighboring subregions give a preferred orientation and a sensitivity to the location and polarity of visual contrast. “Simple” describes the response organization; it does not describe a simpler cell body or a particular shape of neuron.

Hubel and Wiesel's 1962 classification emphasized four observations: distinguishable ON and OFF receptive-field subregions, summation within each region, antagonism between regions, and qualitative prediction of responses to differently shaped stimuli from those regions' arrangement. A slit fitting an elongated center was more effective than a smaller spot; a larger patch could invade opposing flanks and lose effectiveness. An edge separating bright and dark regions could be effective when its boundary matched the boundary between subregions. These were observations about spike trains, rather than an assertion of exact linearity. ^[ref]

The following deliberately coarse signed spatial filter model of a simple cell makes the geometry easy to count. A centered bright bar contributes positively at five sites. Rotating it preserves its area but mixes positive and negative weights; shifting it preserves its orientation but changes which weights it samples. The model is useful because it predicts distinct effects of these three manipulations instead of treating any bright stimulus as equivalent.

A simple cell is not an ideal linear sensor. Its spike train is nonnegative in event count, and neural adaptation can change responsiveness. The signed spatial filter model of a simple cell supplies a useful approximation to spatial drive and an explicit output rule. Comparing predictions with unseen stimuli is more informative than assuming the rule from the name “simple”. In contrast, the existing complex cell models explore tolerance to phase and local displacement.

Spatial summation combines inputs arriving at different synapses. Location matters because a dendrite has axial electrical resistance and leaky membrane capacitance: a local synaptic potential need not arrive at the soma with its original amplitude or time course. A passive dendritic cable makes this attenuation explicit. The superposition principle applies to fixed electric current inputs in a linear passive model, but adding synaptic conductances changes the model's coefficients. Compare voltage changes relative to the same baseline, rather than adding absolute membrane potentials.

A motivating experiment. A 2004 study used $300$–$350\mu \mathrm{m}$ cortical slices from $18$–$28$-day-old rats at $35$–$36{}^{\circ }\mathrm{C}$. Somatic whole-cell recordings measured EPSPs in layer-5 pyramidal neurons. Two fine stimulating electrodes were positioned beside imaged dendrites; calcium-sensitive dye localized activation. Each site was stimulated alone, then both together, usually with paired pulses $20\mathrm{ms}$ apart. Comparing the combined voltage trace with the sum of individual changes tested the summation rule. Nearby inputs on one branch showed strong amplification at intermediate strengths; inputs on different branches were approximately linear.^[ref]

Blocking both excitatory receptor classes eliminated voltage and calcium responses, checking against direct electrical activation. NMDA receptor blockers eliminated or markedly reduced amplification; comparisons with inhibition intact and blocked addressed inhibitory contributions. These controls support local active integration. Passing axons could still recruit unintended synapses; somatic voltage does not measure every local membrane potential. The study tested subthreshold slice responses and pairs of branches, not arbitrary activity in an intact brain.^[ref]

Anatomical links among neural regions. Diffusion-based tractography estimates aspects of white-matter organisation indirectly and should not be equated with a direct count of synapses or information flow.

A synapse is a signaling junction between a neuron and another biological cell. The sending side is presynaptic and the receiving side postsynaptic. At a chemical synapse, released neurotransmitter acts on a receptor in the receiving cell membrane. At an electrical synapse, gap junctions provide a direct path for electric current between cells. See 1.1 Elements of Neuronal Systems.

In a common sequence at a chemical synapse, an arriving action potential opens voltage-gated ion channels selective for calcium. Calcium entry triggers fusion of synaptic vesicles with the presynaptic cell membrane, releasing neurotransmitter into the synaptic cleft. Binding to a receptor changes postsynaptic electrical resistance and conductance, directly through an ion channel or through an intracellular signaling pathway. The resulting change in membrane potential is a synaptic potential. See Ion Channels and the Electrical Properties of Membranes.

A useful conductance approximation writes the inward-positive synaptic electric current aswhere $g_{\mathrm{syn}}\ge 0$ describes the activated electrical resistance and conductance, $V$ is the postsynaptic membrane potential, and $E_{\mathrm{syn}}$ is the reversal potential, the voltage at which this current vanishes. Opening these channels tends to pull $V$ toward $E_{\mathrm{syn}}$; it does not prescribe a current of fixed sign independently of $V$.

For a worked example, let $V=-65\mathrm{mV}$ and $g_{\mathrm{syn}}=1\mathrm{nS}$. A reversal potential of $0\mathrm{mV}$ gives $I_{\mathrm{syn}}=65\mathrm{pA}$, an inward electric current that raises membrane potential. A reversal potential of $-75\mathrm{mV}$ instead gives $-10\mathrm{pA}$, which lowers membrane potential. These are illustrative calculations, not measurements. A reversal potential of $-60\mathrm{mV}$ can raise $V$ from $-65\mathrm{mV}$ yet oppose reaching a threshold of $-50\mathrm{mV}$: a depolarizing input need not be excitatory.

An excitatory synapse favors a subsequent action potential, whereas an inhibitory synapse opposes it. The effect depends on the receptor, the permeating ions, their gradients and the reversal potential; a neurotransmitter does not have one fixed excitatory or inhibitory sign in every context. Even with $E_{\mathrm{syn}}=V$, an added conductance can reduce responses to other inputs by lowering effective electrical resistance, an effect called shunting inhibition. See Purves and colleagues: excitatory and inhibitory postsynaptic potentials. This dependence matters when translating a biological synapse into a signed weight or an imposed electric current in computational neuroscience.

Synaptic conductance describes how strongly activated synaptic ion channels conduct electric current for a given driving force. In the passive approximation, $I_{\mathrm{syn}}=g_{\mathrm{syn}}(E_{\mathrm{syn}}-V)$, with inward-positive electric current. It is the electrical conductance contribution of a synapse; its postsynaptic voltage response also depends on the driving force, the other membrane electrical conductances, and the capacitance. The synaptic potential model and shunting inhibition example show why a nonzero synaptic conductance can have zero electric current at rest.

The proposal that waking-associated strengthening is balanced by sleep-associated renormalization of overall strength of synapses. Selective weakening could improve capacity and relative signal contrast while preserving important differences. It is compatible with some strengthening or reactivation, and is not the claim that sleep indiscriminately erases learning.

Synaptic integration is the conversion of inputs at a neuron's synapses into a changing membrane potential and, in a spiking neuron, the timing or absence of action potentials. Spatial summation concerns inputs at different locations; temporal summation concerns inputs at different times. A useful question is whether individually ineffective inputs can act together, and whether an inhibitory postsynaptic potential changes that cooperation. The answer depends on location, timing and the receiving neuron's state.

A synaptic potential is a voltage response, whereas synaptic conductance describes activated ion channels. With inward-positive electric current, the contribution of input $j$ isHere $g_j$ is its synaptic conductance, $E_j$ its reversal potential and $V$ the local membrane potential. The driving force $E_j-V$ is an electric potential difference. An EPSP generally brings the membrane potential toward a more positive reversal potential, but its driving force shrinks during depolarization. An IPSP need not produce hyperpolarization: a reversal potential above rest but below the neuronal firing threshold can produce depolarization while opposing a larger excursion. At $E_j=V$, the added synaptic conductance initially carries zero electric current yet can cause shunting inhibition of another input.^[ref]

Two distinct nonlinearities matter. Even a passive cell membrane can show sublinear synaptic integration when added synaptic conductances reduce driving force and effective electrical resistance. Active voltage-gated ion channels and NMDA receptors can instead amplify sufficiently concentrated inputs and produce a dendritic spike. Neither summing prescribed electric currents nor adding separate measured EPSPs is an exact universal rule for a branching neuron.^[ref]