Codex @codex  0

A neuronal process that carries signals toward output sites, often by propagation of action potentials to synapses.

A family of models in which an instructive signal changes the influence of recently active inputs on cerebellar output. In a smooth-pursuit experiment, complex-spike duration predicted subsequent simple-spike depression and eye-movement learning; matching preceding simple-spike activity provided a control. This supports graded instruction during behaviour without proving an exact delta rule at every synapse.^[ref]

A chemical synapse transmits a signal by releasing neurotransmitter from a presynaptic neuron to act on postsynaptic receptors across a synaptic cleft. The two cell membranes remain separate. A typical fast transmission sequence is:

See Chemical Synapses and 3.1 Synapses. The fast pathway uses synaptic vesicles prepared near release sites; it does not require a distant synaptic vesicle to travel across the terminal after every action potential. See Basic Neurochemistry: calcium entry and vesicles at release sites. Some chemical synapses release neurotransmitter in response to graded presynaptic membrane potential changes rather than action potentials.

Neurotransmitter clearance limits how long released neurotransmitter remains available. Depending on the neurotransmitter and location, diffusion away, uptake through membrane transport proteins into the sending or neighboring biological cells, and breakdown by enzymes contribute. Receptor binding and unbinding also affect response duration. Recovering synaptic vesicle membrane is a separate recycling task; uptake of released neurotransmitter is not the return of an intact synaptic vesicle. See General Principles of Cell Communication.

This electrical-to-chemical-to-electrical sequence offers several places to adjust a signal: synaptic release probability, receptor availability and neurotransmitter clearance can change its strength or duration. It helps explain why synaptic plasticity can modify communication between neurons, and why an arriving action potential need not cause a postsynaptic action potential.

Otto Loewi's 1936 Nobel lecture describes his 1921 experiment: fluid from a frog heart whose nerves had been stimulated produced the corresponding response when transferred to another heart. The transferable substance established chemical mediation in that preparation. This was evidence for released chemical signaling, not yet an identification of synaptic vesicle fusion and not a claim that every synapse is chemical. See Loewi, The Chemical Transmission of Nerve Action, Nobel lecture, 12 December 1936.

Extracellular single-unit recording uses an electrode outside a neuron to detect electrical events attributed to that unit's action potentials. It records a spike train, not the intracellular membrane potential or every input synapse. ^[ref]

The leaky integrate-and-fire model combines a passive equation for membrane potential with a firing rule. Between events, inward-positive applied electric current $I(t)$ obeyswhere $C>0$ is membrane capacitance, $g_L>0$ is the leak conductance, $R$ is the corresponding electrical resistance, and $E_L$ is the leak reversal potential. Here the leak conductance is constant. The time constant $\tau$ describes passive relaxation. When $V$ first reaches a threshold $V_{\theta }$ from below, record an action potential event and reset $V$ to $V_r<V_{\theta }$. In the version used here, hold $V=V_r$ for a fixed neuronal refractory period $t_{\mathrm{ref}}$, then resume integration. See 1.3 Integrate-And-Fire Models.

For constant $I$, with $t=0$ at the start of an integration interval and $V(0)=V_{\mathrm{initial}}$, the exact subthreshold solution isThe equilibrium point of this passive equation is $V_{\infty }=E_L+RI$. Subtracting $V_{\infty }$ leaves $d(V-V_{\infty })/dt=-(V-V_{\infty })/\tau$, explaining the exponential function. This solution applies only until a threshold event or a change in $I$. Starting below threshold, constant input produces repeated events if $V_{\infty }>V_{\theta }$; if $V_{\infty }=V_{\theta }$, the threshold is approached asymptotically and never reached at finite time.

For a worked example, chooseThese imply $g_L=10\mathrm{nS}$ and $C=200\mathrm{pF}$. All these parameters are illustrative model choices, not measurements from a particular neuron. With $I=200\mathrm{pA}$, Ohm's law gives $RI=20\mathrm{mV}$ and hence $V_{\infty }=-45\mathrm{mV}$. ThusAt threshold, $e^{-t/\tau }=1/4$, so the first event occurs after $T=20\ln 4\mathrm{ms}\approx 27.7259\mathrm{ms}$. Each subsequent cycle includes the $2\mathrm{ms}$ neuronal refractory period, givingThis is the ideal sustained firing rate; the first-event latency is $T$, not $T+t_{\mathrm{ref}}$.

With $I=100\mathrm{pA}$ instead, $RI=10\mathrm{mV}$ and $V_{\infty }=-55\mathrm{mV}$. From the same initial state, $V(t)=-55\mathrm{mV}-10\mathrm{mV}{e}^{-t/\tau }$ stays below $-50\mathrm{mV}$: no events occur, however long this current is held. The critical sustained current is $(V_{\theta }-E_L)/R=150\mathrm{pA}$; equality gives asymptotic approach rather than finite-time firing.

The leak gives old inputs progressively less influence, while capacitance prevents an ordinary finite electric current step from changing $V$ instantly. This combination makes the leaky integrate-and-fire model useful for exploring input summation, event timing and network activity at low computational cost. Its fixed parameters and reset omit the evolving voltage-gated ion channels of the Hodgkin-Huxley model, spatial integration in dendrites and mechanisms of neural adaptation. Treating a synapse as a prescribed $I(t)$ also omits its voltage-dependent driving force unless a conductance term is added. See 1.4 Limitations of the Leaky Integrate-and-Fire Model and constant-input firing rate of a leaky integrate-and-fire neuron.

A neuron is a specialized biological cell that receives, integrates and transmits signals. Its shape separates places where inputs arrive from places where outputs reach other biological cells: this functional asymmetry is neuronal polarity. Understanding a neuron therefore means connecting its anatomy to changes in membrane potential, rather than treating it as a wire that merely forwards every input.

In a common multipolar neuron, branching dendrites receive inputs at synapses; the soma contains the cell nucleus and most major organelles; the axon carries signals to distant or nearby targets; and axon terminals form output contacts. All these regions belong to one biological cell enclosed by a continuous cell membrane. An axon can branch, allowing one neuron to influence several receiving biological cells.^[ref]

At an input chemical synapse, a neurotransmitter acts on a receptor, changing an electric current directly or through signal transduction.^[ref] A resulting synaptic potential contributes to synaptic integration in the dendrites and soma. In many spiking neurons, sufficiently strong depolarization recruits voltage-gated ion channels in the axon initial segment, initiating an action potential that propagates along the axon. At an output chemical synapse, calcium entry through voltage-gated ion channels promotes synaptic vesicle exocytosis, releasing neurotransmitter. An action potential is thus regenerated electrically along a cell membrane; it is not a packet of neurotransmitter travelling from the dendrites to the axon terminals.^[ref]

The compartments are specializations, not exclusive jobs. Inputs can contact the soma or axon as well as dendrites; dendrites can support active voltage-gated ion channels. Paired recordings from a pyramidal neuron's soma and dendrite or axon established that an action potential can begin in the axon and propagate back into the dendrites. The usual output direction therefore does not make every electrical signal strictly one-way.^[ref] An electrical synapse, which directly couples cytoplasms through gap junctions, also differs from the sending-and-receiving arrangement of a chemical synapse.^[ref]

Nor does every neuron have the upper diagram's shape. A bipolar neuron has two main processes, one dendrite and one axon; a multipolar neuron has several dendrites; and a pseudounipolar neuron has a single process from the soma that divides into peripheral and central branches. In the last arrangement, sensory signals can travel past the soma rather than obligatorily through it. These shape categories do not specify every neuron's receptor, connectivity or firing behavior.^[ref] Some neurons in the retina communicate through graded changes in membrane potential without generating action potentials.^[ref]

Why timing changes the computation. Synaptic integration can combine inputs at different locations, called spatial summation, and inputs arriving at different times, called temporal summation. A minimal example uses the leaky integrate-and-fire model with the illustrative parameters already given there: resting and reset membrane potential $-65\mathrm{mV}$, firing threshold $-50\mathrm{mV}$, electrical resistance $R=100\mathrm{M\Omega }$, capacitance $C=200\mathrm{pF}$, time constant $\tau =RC=20\mathrm{ms}$ and neuronal refractory period $2\mathrm{ms}$. These are chosen model values, not measured properties of a particular neuron. Before firing, a constant inward electric current $I$ starting from rest givesThe exponential function expresses the competition between charging the cell membrane's capacitance and leakage through its electrical resistance.^[ref]

Two $100\mathrm{pA}$ inputs, each lasting $40\mathrm{ms}$, produce different outputs depending on their timing. Either input alone reaches only $-56.3534\mathrm{mV}$. Applied together, they give $I=200\mathrm{pA}$ and reach threshold after $20\ln 4\mathrm{ms}\approx 27.7259\mathrm{ms}$, producing one action potential event during the pulse. If instead the second input begins $20\mathrm{ms}$ after the first ends, the earlier depolarization has partly decayed: the second pulse peaks at approximately $-55.9229\mathrm{mV}$ and neither pulse produces an action potential. The two-input protocols deliver the same total electric charge, $8\mathrm{pC}$, yet only the coincident protocol fires. Counting total input alone misses this timing dependence.

This example makes a neuron a detector of sufficiently clustered input, not a perfect logical gate: different parameters change the result. Actual dendrites introduce location-dependent filtering and active responses, and an inhibitory postsynaptic potential or shunting inhibition can change whether another input succeeds. The single-compartment leaky integrate-and-fire model omits these mechanisms.^[ref] Across connected neurons, branching axons distribute outputs while converging synapses combine inputs; spike trains can support temporal coding and rate coding, and synaptic plasticity changes the influence of an input. Connectivity and cellular dynamics together determine the computation.^[ref][ref]

Testing the connection between structure and function. Whole-cell recording connects anatomy to a measured membrane potential. In a published mouse protocol, a two-photon microscope guides solution-filled glass pipettes to neurons; an amplifier and bath reference electrode complete the electrical recording apparatus. A high-resistance seal forms against the cell membrane, then brief suction opens the membrane patch beneath the tip. In current clamp, the amplifier supplies a chosen electric current and records the resulting membrane potential, unlike voltage clamp, which controls voltage and measures electric current. The investigators used $500\mathrm{ms}$ current steps of several amplitudes to characterize excitability and firing, and negative current pulses to estimate input electrical resistance.^[ref]

The observable is a voltage trace, including graded responses and action potentials. Comparing injected electric current with that trace tests how input becomes output; it does not identify every underlying ion channel. The protocol checks access electrical resistance, which can distort voltage and action potential measurements, and includes intervals without stimulation to record spontaneous activity. In a teaching replication, repeated zero-input trials and monitoring baseline drift would provide further controls; these are recommended comparisons, not additional results attributed to the paper. A recording at the soma also cannot establish the local voltage everywhere in a branching dendrite, and this anesthetized preparation does not establish behavior in every awake neuron.^[ref]

A complementary experiment used simultaneous whole-cell recordings from the soma and an apical dendrite or axon of the same pyramidal neuron in a brain slice. The relative onset times of action potentials, together with sodium currents in excised dendrite membrane patches, supported axonal initiation followed by active propagation into dendrites. Pairing sites within one neuron is a useful comparison because it avoids confusing differences between cells with differences between compartments. Instrument timing and comparable recording quality would be essential controls when reproducing the inference. Only the published abstract was accessible here, so no unverified timing values, drug controls or channel distributions are supplied.^[ref]

Neuronal paired-pulse recovery probes the effect of an earlier action potential with a later test stimulus. First establish a conditioning action potential; then vary the delay $\Delta$ to a test pulse of specified amplitude, duration and delivery site. Record whether a second action potential occurs, its timing and, where relevant, its waveform. Repeating the test across amplitudes distinguishes recovery of responsiveness to one pulse from recovery of the minimum effective pulse. Direct current clamp stimulation can probe the receiving neuron's excitability; stimulation through a synapse additionally allows synaptic plasticity to alter the input.

A documented whole-nerve teaching experiment keeps a bundle of axons in physiological solution, with its proximal end drawn into a stimulating tube and its distal end into a recording tube. The suction electrodes use short plastic tubes and $1\mathrm{ml}$ syringe-plunger terminals, held by O-rings. Longitudinal electric currents create a voltage drop in the thin solution layer between bundle and tube, read by a cathode-ray oscilloscope. Electrode fit affects amplitude. An immediate stimulation artifact remains even without the nerve, whereas the neural signal arrives after a delay.^[ref][ref]

The paired-stimulus exercise sets $0.1\mathrm{ms}$ pulses to obtain a maximal undistorted A wave of the compound action potential, then increases the spacing. The described first response remains constant while the second grows with spacing. Different axons recover at different times; the summed response therefore grows as more contribute. A smaller second compound action potential does not mean every individual action potential has shrunk. The source contrasts this with a single axon's response or failure. Its response-amplitude recovery criterion is not a universal single-cell threshold criterion.^[ref]

Useful additional controls would include single-pulse baselines interleaved with paired trials, a verified conditioning action potential, enough rest between trials, and fixed temperature, electrode geometry, pulse width and initial membrane potential. Confirm stimulus delivery and amplifier recovery, separate the stimulus artifact from a second neural event, and check whether waveforms overlap before measuring amplitudes. These are proposed checks, not additional results claimed for that teaching experiment. Even a clean compound action potential does not uniquely identify sodium availability: recruitment, propagation and synchronization of different axons influence the summed signal. Nor does failure at a distant recording site alone distinguish failed initiation from failed propagation.

ON and OFF receptive-field subregions distinguish locations by response polarity. In the classic mapping experiment, a bright spot in an ON region increased firing during illumination; illumination in an OFF region suppressed maintained firing and/or evoked a discharge when the light disappeared. If maintained firing was very low, suppression could be hard to see. The recorded decrease alone did not distinguish direct inhibition at a synapse from reduced excitation. ^[ref] These are functional regions in visual space, not compartments of a cell body.

A branch of a cerebellar granule-cell axon that forms synapses on Purkinje cells and other cerebellar neurons. Expanded granule-cell activity supplies a diverse input basis in cerebellar adaptive-filter models.

Sleep can protect newly acquired information, actively reorganize it, and restore the capacity for subsequent learning. These are distinct claims. Memory consolidation concerns changes after acquisition; better learning the next morning concerns encoding capacity; less interference during an unoccupied interval need not imply an active sleep mechanism. The strongest explanation therefore combines behavioural comparisons with recordings and selective interventions.

A basic retention experiment compares later performance after sleep and waking, ideally relative to an immediate baseline. Yet a night-versus-day comparison also changes circadian phase, interference and alertness at testing. Sleep deprivation adds physiological stress and attentional impairment. Matching acquisition, testing at comparable times, measuring vigilance, and comparing sleep with quiet waking make the inference sharper. A retention advantage after sleep remains compatible with both passive protection and active processing; distinguishing them requires manipulating what happens during the interval.

The active systems consolidation hypothesis provides an active mechanism. Rapid relational encoding by the hippocampus permits later memory reactivation to modify distributed cerebral cortex representations. During NREM sleep, cortical sleep slow oscillations can organize sleep spindles and hippocampal sharp-wave ripples, creating coordinated windows for synaptic plasticity. This is a timing hypothesis about communication, not proof that a complete memory is copied from one anatomical container to another. Systems consolidation and synaptic consolidation can proceed together, and some detailed episodic memories may continue to require hippocampal contributions.

An important recording experiment monitored ensembles of hippocampal place cells in rats during spatial behaviour and during sleep before and after it. Cells active together at particular locations became more likely to fire together in subsequent slow-wave sleep; the pre-experience recording supplied a valuable baseline. This demonstrates experience-related memory reactivation. Pairwise coactivation alone does not establish ordered neural replay, nor does a correlation show that the events improve retention. An apparently memory-related pattern might otherwise be an incidental consequence of prior activity.^[ref]

A stronger causal test interrupted hippocampal sharp-wave ripples during post-training sleep in rats learning a spatial task. An online electrical recording system detected ripple events and triggered brief disruption; a delayed-stimulation comparison helped control stimulation without abolishing the same events. Spatial performance was impaired, while gross sleep architecture was not detectably changed. This supports a functional role for ripple-associated activity beyond simply obtaining sleep. Nevertheless, disruption affects an event and its associated activity, not an experimentally isolated memory message; it does not identify the exact information transferred or establish that all memory consolidation requires the same mechanism.^[ref]

Human targeted memory reactivation offers complementary evidence. Participants learned object locations with an odour present, then received that odour again during polysomnographically identified slow-wave sleep. Retention improved compared with an odourless condition. Crucially, the benefit did not appear when the odour had been absent during learning, or when it was presented during waking or REM sleep; procedural finger-tapping did not show the same benefit. The cue–learning relationship and sleep state therefore mattered. The associated hippocampal activation is consistent with memory reactivation, although activation does not reveal all the represented content. This is stronger evidence for selective processing than an undifferentiated sleep/wake advantage.^[ref]

Sleep could also benefit memory by regulating the substrate for learning. The synaptic homeostasis hypothesis proposes that predominantly waking-associated strengthening requires compensatory renormalization, limiting saturation, energy costs and interference. Selective preservation can improve the contrast of important connections even if average strength of synapses falls. Serial electron-microscopy reconstructions of mouse cortical synapses found smaller axon–spine interfaces after sleep, with an average decrease of roughly 18%; the largest synapses were relatively spared. The comparison included spontaneous and enforced waking. These are structural measurements consistent with selective scaling, not a direct measurement of every synapse's efficacy or proof that the measured changes caused a particular memory benefit.^[ref]

Renormalization and selective strengthening are not mutually exclusive. A learning-related ensemble could be selectively stabilized while less useful background connections weaken. Memory reactivation could help select what survives, and improved capacity after sleep could support the next episode rather than merely preserve the last one. An adequate theory must explain both specificity and restoration, rather than equate more firing or larger synapses with better memory everywhere.

The interpretation must remain task-specific. A cueing effect for declarative memory does not assign every procedural memory to REM sleep, and a null cueing result in one stage does not show that stage has no mnemonic function. Sleep-stage manipulations can alter other stages or awaken subjects. Rodent spatial performance provides mechanistic access but does not by itself establish human autobiographical experience. Replicated behavioural effects, adequate controls and selective circuit perturbations are most persuasive when they converge.

Thus sleep is important because it provides a state in which information can be selectively re-expressed and reorganized while conditions at synapses are regulated. Evidence available by April 2022 supports active contributions in addition to passive protection; it does not establish a universal sleep-only route for every kind of learning.

Monoamine signalling contributes to mood regulation and to the actions of several antidepressants, but a universal “low transmitter causes depression” account is not established. Three propositions must be separated: a drug changes a molecular target; that change helps produce improvement; and the untreated disorder was caused by an opposite abnormality. Evidence for one proposition does not logically establish the next.

The monoamine hypothesis of depression developed from pharmacology. Monoamine oxidase inhibition reduces transmitter metabolism, while some tricyclic antidepressants inhibit transmitter reuptake; SSRI principally inhibit serotonin reuptake. Norepinephrine, serotonin and dopamine are distinct monoamines, acting through different receptors and pathways. The historical catecholamine proposal emphasized norepinephrine and already acknowledged the limitations of extrapolating from drug effects to disease mechanisms. Treating every low affective state as depletion of a single interchangeable brain substance is a substantially stronger claim.^[ref]

Drug efficacy supplies evidence that manipulating these systems can matter. A 2018 network meta-analysis synthesized 522 double-blind trials in adults with major depressive disorder and found each included antidepressant more efficacious than placebo on its response outcome. This is evidence synthesis, not a direct assay of patients' transmitter levels. Average effects, trial selection, publication bias, blinding and symptom-scale outcomes limit interpretation, and not every patient benefits. Even a perfectly demonstrated effect would show that an intervention can change a pathological state, not that it reverses its original cause. A downstream compensatory mechanism can be therapeutically useful without being the initiating lesion.^[ref]

Measuring the proposed deficiency is itself difficult. Peripheral serotonin is not a direct measurement of release at a particular central synapse. Metabolites combine synthesis, release, reuptake and degradation; receptor or transporter imaging is not simply a concentration gauge. A difference observed after illness onset might be causal, compensatory, treatment-related or associated with physiological stress. Establishing a general deficiency would require convergent measures in appropriate untreated populations, temporal precedence and a relationship to well-defined symptoms, not merely one group-average biomarker.

Precursor depletion supplies a more informative intervention. In a double-blind crossover study, patients in antidepressant-induced remission underwent acute tryptophan depletion and a control condition. Fourteen of 21 experienced a transient depressive relapse after depletion, whereas control testing did not produce significant behavioural effects. This supports dependence of some maintained remissions on serotonin precursor availability. It does not show that all untreated depressed patients began with low serotonin, that depletion is sufficient to generate the full syndrome in everyone, or that norepinephrine and dopamine are irrelevant. The selected, previously treated sample is central to the causal question actually tested.^[ref]

Time course also challenges a simple replenishment story. Initial transporter inhibition can occur before substantial clinical improvement. That discrepancy motivates slower receptor, circuit, learning and plasticity changes; it does not imply that early drug effects are absent. In a randomized placebo-controlled experiment, an acute dose of a norepinephrine reuptake inhibitor changed positive emotional processing in depressed participants before subjective mood or anxiety changed. Facial-expression judgments and emotionally valenced categorization and recall supplied behavioural readouts. These observations support an affective processing bias account: altered experience processing could accumulate into mood change. They do not establish that the measured early shift mediates every later clinical response.^[ref]

Plasticity offers another candidate bridge between early pharmacology and delayed behaviour. In mice, preventing antidepressant-associated adult hippocampal neurogenesis blocked behavioural effects in particular experimental assays. The evidence included genetic and localized irradiation approaches. It supplies a mechanistic dependency within that model, rather than merely a correlation between a drug and cell proliferation. However, the assays are not human major depressive disorder, the interventions can have additional effects, and necessity for one behavioural endpoint does not establish either sufficiency or a universal human mechanism.^[ref]

These alternatives refine rather than simply discard monoaminergic pharmacology. Serotonin or norepinephrine could regulate plasticity and learning within distributed neural circuits; baseline concentrations need not be uniformly low. Receptor sensitivity, pathway-specific activity and environmental experience could determine the effect of an intervention. Conversely, observing altered dopamine-related motivation would not prove that all anhedonia, cognitive slowing and sadness have one cause. Major depressive disorder groups multiple symptom patterns, so a mechanism associated with one dimension can be diluted or misleading when treated as a universal disease explanation.

The strongest conclusion available at the examination date is therefore conditional. Particular monoamine manipulations can alter affective processing and symptoms, and some treatment responses depend on monoaminergic function. The evidence does not identify a single brain-wide deficit that is necessary and sufficient for human depression. A useful successor account specifies transmitter, receptor, circuit, time scale and symptom, and tests how biological changes interact with learning and environment. This evaluation concerns explanatory evidence, not a recommendation to start, stop or choose a medication.

Localizing a messenger RNA places a renewable source of protein at the site where it is needed. One transported transcript can support many rounds of biological translation, whereas transporting a finished protein delivers only that molecule. Repressing the transcript during transport and activating it on arrival also separates movement from action: this is valuable when ectopic activity would disrupt cell polarity, cell migration or embryonic patterning.

The advantage is especially clear in neurons. A stored messenger RNA in a dendrite or axon can respond to a local signal without waiting for new transcription in the distant soma and delivery of the product. Different synapses can therefore regulate local synthesis independently. Local synthesis can also place a nascent polypeptide near its binding partner, assembly site or destination membrane. Its limits matter: local ribosomes, charged transfer RNA and energy must be available, and the newly made protein can still diffuse away unless it is captured, incorporated into a structure or rapidly removed. Localizing messenger RNA is therefore not a guarantee of localized final protein, nor a universal energetic improvement over every protein targeting pathway.

Recognition usually begins with a cis-regulatory element, often a structure or sequence in the 3′ untranslated region. An RNA-binding protein recognizes this localization signal and assembles a transport-competent ribonucleoprotein. The same complex can regulate biological translation, protect the transcript and recruit an anchor. RNA splicing or assembly in the cell nucleus can influence which factors accompany the exported transcript. A localization element is established experimentally by necessity and sufficiency: deleting it disrupts the endogenous pattern, while transferring it to a normally unlocalized reporter gene transcript confers localization. Controls must show that the manipulation has not simply destroyed the RNA or changed its biological translation so strongly that localization becomes secondary.

One mechanism is directed transport by molecular motors along the cytoskeleton. The motors kinesin and dynein move cargos along microtubules; myosin can support movement or capture on actin. Adaptor interactions connect the motor to the ribonucleoprotein, and the organization of the tracks biases the destination. Individual particles can move in both directions, pause and reverse. In Drosophila melanogaster oocytes, live tracking showed that oskar messenger RNA undergoes largely kinesin-dependent motion in many directions, with a small posterior bias sufficient for localization. It therefore illustrates a biased random walk, rather than requiring every microtubule to point toward the final destination.^[ref]

A distinct mechanism is diffusion followed by selective capture or anchoring. Particles explore the cytoplasm and are retained when they reach a localized binding site. Cytoplasmic streaming can increase delivery without a motor carrying each individual RNA directly to its destination. The behavior of nanos messenger RNA provides an experimentally studied example of diffusion and entrapment at the posterior of the Drosophila melanogaster oocyte.^[ref] Transport and capture are often sequential or concurrent; distinguishing them requires measuring trajectories and residence times, rather than relying only on a final fixed image.

A third mechanism is spatially selective stability. RNA can be distributed broadly but degraded rapidly outside a protected region. The localized population then reflects longer residence through slower decay, even if delivery itself is unbiased. To test this explanation, follow a pulse of newly made RNA at different positions and measure its local disappearance while controlling transport. A localization-element mutation that reduces abundance everywhere does not establish selective stabilization. Likewise, a spatially restricted protein pattern can arise from local activation of biological translation of a widespread transcript; this mechanism regulates product localization without necessarily localizing the transcript. Measuring both species distinguishes these possibilities.

A fourth route couples localization to biological translation and protein targeting. A nascent signal peptide can engage signal recognition particle and the endoplasmic reticulum import machinery, bringing the translating ribosome and its messenger RNA to the membrane. Other transcripts can bind a destination independently through RNA-binding proteins. Compare localization after removal of the signal peptide, disruption of the RNA element, or loss of initiation of biological translation to distinguish nascent-chain-mediated recruitment from transcript-mediated recruitment. Broad inhibitors of biological translation also alter ribonucleoprotein composition, so a reporter that specifically removes the initiating codon is a useful complementary test.

Localization must be coordinated with activation. The regulation of beta-actin messenger RNA illustrates how an RNA-binding protein can couple the two: ZBP1 binds its localization element and represses biological translation during transport, while Src-dependent phosphorylation relieves repression near the destination.^[ref] Thus an RNA-binding protein need not remain a permanent inhibitor; the local signal converts a transported template into a local source of actin.

Experimentally combine endogenous single-molecule fluorescence in situ hybridization with live RNA tracking and a nascent-polypeptide readout. Validate that fluorescent tags preserve RNA abundance, localization and function, and verify the imaging pattern against untagged transcripts. Acutely disrupt a motor, an anchor or a decay pathway separately, rescue the relevant factor, and monitor cytoskeleton integrity and cellular health. Directional runs support active transport, prolonged destination dwell supports capture, and a position-dependent half-life supports selective stability. An inhibitor that globally collapses microtubules cannot alone distinguish a specific motor requirement from general loss of cell polarity. The persuasive explanation connects delivery, retention and local synthesis to the measured biological outcome.

A two-layer heteroassociative memory can retrieve one activity pattern from a different associated cue through learned feedforward synapses. Let $x^{\mu }\in {\mathbb{R}}^n$ represent a cue and $t^{\mu }\in {\mathbb{R}}^m$ its associate. An input population projects to an output population through $W$, with retrieval $y=\varphi (Wx-\theta )$. The output activation function can be linear for an analysis or thresholded with synaptic competition for a sparse categorical response. This is pattern association: the output need not reproduce the input, unlike an autoassociative memory.

During storage, the cue and its desired associate must be coactive. A local Hebbian learning rule, $\Delta W_{ij}=\eta t_i{x}_j$, strengthens the connection from an active cue neuron to an active target neuron. With normalized inputs, the idealized batch weights arePresenting cue $x^{\nu }$ givesThe first term is the desired association; the second is interference from overlapping cues. Thus retrieval follows exactly for orthonormal cue patterns, while sparse coding, centring, synaptic normalization and a competitive output can reduce interference for realistic overlapping inputs. These devices have distinct roles: synaptic normalization prevents unlimited weight growth, and synaptic competition separates competing associates. Signed, centred model weights should not be identified literally with negative excitatory synapses; biological inhibition and baseline subtraction need an explicit circuit implementation.

As an exact illustrative calculation, take $x^1=(1,0{)}^{\mathsf{T}}$, $x^2=(0,1{)}^{\mathsf{T}}$, $t^1=(1,1,0{)}^{\mathsf{T}}$ and $t^2=(0,0,1{)}^{\mathsf{T}}$. ThenA cue $(0.8,0.1{)}^{\mathsf{T}}$ produces $(0.8,0.8,0.1{)}^{\mathsf{T}}$, so threshold $0.5$ retrieves the first associate. These are chosen model values, not recorded firing rates. If two genuinely different associates share an identical cue, a deterministic feedforward map cannot choose between them without additional context; adding capacity alone does not solve that ambiguity. A cue-to-associate map also need not provide the reverse association unless the reverse pathway is separately learned.

The hippocampus provides a plausible anatomical setting. In a hippocampal episodic memory model, a retrieved hippocampal CA3 event pattern can project through the hippocampal CA1 output pathway to reinstate associated cortical content. The feedforward CA3-to-CA1 and hippocampal-to-cortical links can provide heteroassociation, while recurrent CA3 connectivity provides autoassociation. The entorhinal cortex supplies converging information and routes cortical communication. This is a computational assignment to pathways; anatomical connectivity alone does not prove that their weights implement the displayed rule.

An especially direct physiological test comes from trained visual pairs. Monkeys learned arbitrary associations between computer-generated pictures and then selected the associated picture after a delay. Extracellular single-unit recording in anterior temporal cortex found both neurons responding to the two members of a learned pair and neurons whose delay activity anticipated the appropriate associate. The preparation therefore links an experimentally controlled cue–associate relation to neural responses, rather than merely showing visual selectivity. However, the recordings do not identify the individual modified synapses, establish one-way connectivity, or exclude recurrent retrieval. They support an associative representation in temporal cortex, not a unique proof of a two-layer implementation.^[ref]

The rapid-learning requirement is relevant to the hippocampal implementation: CA3-restricted NMDA receptor deletion impaired learning of a newly located hidden platform in a delayed matching-to-place task while sparing performance with previously experienced locations. This supports rapid hippocampal storage, but does not by itself identify a specific feedforward association matrix.^[ref]

A decisive model test would measure cue and associate representations before learning, after learning, and during partial-cue retrieval, while selectively disrupting the proposed connecting pathway. Correct associate decoding with unrelated-cue and familiar-visual-response controls would be stronger evidence than activity during successful recall alone. The architecture is a plausible component of temporal and hippocampal memory circuits; its local learning rule, interference properties and pathway-specific causal predictions make it testable.

A useful hippocampal episodic memory model assigns complementary computations to competitive learning, recurrent autoassociation and feedforward pattern association. The event must bind its participants, location and context, so that a partial cue later reinstates the appropriate conjunction. Merely classifying inputs or remembering a location is insufficient to explain an episodic memory.

A concrete pathway begins with distributed cortical input through the entorhinal cortex. The dentate gyrus expands and sparsifies this input: strong synaptic competition means that only a small fraction of granule cells respond strongly. A winning unit or winning sparse population strengthens its active afferents through constrained Hebbian learning. Similar episodes can then recruit less-overlapping outputs, providing pattern separation before storage. Separating two nearly identical episodes is useful because a recurrent memory otherwise tends to merge their cues. The desired separation must preserve relevant conjunctions; indiscriminately randomizing the input would make later retrieval impossible.

The sparse code recruits a hippocampal CA3 ensemble. Rapid, associatively induced long-term potentiation is a candidate substrate for storing the conjunction, with its induction conditions and persistence requiring experimental constraints. Its recurrent collateral synapses allow coactive elements of the event to become associated through Hebbian learning. One idealized centred storage rule iswhere $a$ is mean coding activity and ${\xi }^{\mu }$ is an event pattern. With inhibitory control maintaining sparsity, recurrent retrieval iterates $r\leftarrow \varphi (Wr+c-\theta )$ from cue $c$. A sufficiently informative partial cue can move activity into the stored event's basin of attraction; recurrent reinstatement supplies missing components. This is pattern completion. A Hopfield network makes the principle precise in a symmetric idealization with an energy landscape, but biological CA3 connectivity need not obey the exact symmetry or update assumptions of asynchronous Hopfield energy descent. Interference, spurious states and an incorrect initial cue can all produce erroneous completion.

The hippocampal CA1 and cortical return pathways provide heteroassociative memory. During the event, CA3 activity ${\xi }^{\mu }$ is paired with a cortical or CA1 target $z^{\mu }$, giving an illustrative mapping $A={\sum }_{\mu }{z}^{\mu }({\xi }^{\mu }{)}^{\mathsf{T}}$. When CA3 completes the event, $A\hat{\xi }$ reinstates its associated output representation. Thus competitive learning distinguishes inputs, autoassociation restores a stored conjunction, and pattern association translates the retrieved conjunction into content usable elsewhere. These are separable operations even if a real circuit performs them concurrently. CA1 also receives direct entorhinal information, so it cannot be treated as a passive relay without testing the contribution of that second input.

There is mechanistically relevant evidence beyond anatomical resemblance. In adult mice, CA3-restricted NMDA receptor deletion spared ordinary acquisition and retrieval of water-maze reference memory with the full cue set, but impaired retrieval when only some cues remained. CA1 place-related activity also deteriorated under partial-cue conditions. Full-cue performance is an important control against interpreting the deficit as a general inability to swim, see cues or express a learned response. This fits the prediction that recurrent associative plasticity is particularly important for pattern completion. Nevertheless, the deletion affects receptor-dependent processing throughout the targeted CA3 cells, and does not isolate only recurrent collateral plasticity or the exact storage rule.^[ref]

A separate delayed matching-to-place experiment found impaired rapid acquisition of novel platform locations after CA3-restricted NMDA receptor deletion, with previously experienced locations relatively spared. That extends the interpretation from partial-cue recall to rapid storage of new experiences; it also cautions against assigning CA3 an exclusively retrieval-related role.^[ref]

Simultaneous dentate gyrus and hippocampal CA3 recordings offer a more direct input–output comparison. In behaving rats, conflict between local and global spatial reference frames disrupted dentate population patterns, while CA3 maintained a more coherent representation related to the stored map. Comparing the two populations under the same manipulation supports the complementary pattern separation and pattern completion computations more specifically than finding different firing rates in different areas. It still does not prove that a particular recalled human episode corresponds to one fixed CA3 attractor.^[ref]

The evidence supports the computational division of labour, but the simplest implementation is incomplete. The same region can separate strongly changed inputs and complete moderately degraded ones; separation and completion are operating regimes, not immutable labels for every trial. Fast storage must coexist with protection of old memories, and recurrent completion must not overwrite a new but similar event. Novelty, inhibition and synaptic plasticity regulation are therefore functional requirements, rather than arbitrary additions. Experimental interventions can affect acquisition, retrieval, exploration and sensory input differently, so their timing and behavioural controls matter.

An unordered fixed-point memory also omits the temporal structure of an episode. The same people in the same place at different times need contextual or sequential representations; an attractor for a bag of features does not distinguish reversed event order. Rehearsal and longer-term cortical learning may change the dependence on the hippocampus. The theory should consequently predict partial-cue success, lure errors, interference and pathway-specific deficits across time, while acknowledging that neither a spatial task nor a correlation in population activity establishes the whole of episodic memory. The combined architecture supplies an explicit storage-and-retrieval mechanism, with experimentally supported components and clear remaining requirements.

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.