Associative learning enhances population coding by inverting inter-neuronal correlation patterns

Associative learning enhances population coding by inverting inter-neuronal correlation patterns Jeanne Sharpee Gentner

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

Associative learning enhances population coding by inverting inter-neuronal correlation patterns 2013 Neuron 10.1016/j.neuron.2013.02.023 Jeanne2013-associative-learning-enhances-population-cod Jeanne2013-associative-learning-enhances-population-cod.txt, the project's staged copy bf1b3077509068100615c18feab80157556443d2 not-redistributable neuroscience

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them how associative learning changes the pattern of correlations between neurons, and what that does to population coding Quotation withheld because this record is not cleared for redistribution. whether learning acts on emergent population properties that cannot be observed in single neurons Quotation withheld because this record is not cleared for redistribution. shows learning inverts the relationship between signal and noise correlations, but only for stimuli that guide behaviour, and that this improves population decoding increasingly with population size Quotation withheld because this record is not cleared for redistribution. mechanism Quotation withheld because this record is not cleared for redistribution. inter-neuronal noise correlations in population coding Quotation withheld because this record is not cleared for redistribution. electrophysiology electrophysiology Quotation withheld because this record is not cleared for redistribution. mechanisms when noise correlation runs opposite to tuning similarity, a downstream reader can subtract the responses of dissimilarly tuned neurons, cancelling the shared noise while adding the signal, so the same spikes carry more information Quotation withheld because this record is not cleared for redistribution. mechanisms learning-driven changes to feed-forward synaptic weights (with excitatory-inhibitory noise cancellation) decorrelate similarly tuned neurons and preserve or increase correlation among dissimilarly tuned ones, producing the inverted correlation structure Quotation withheld because this record is not cleared for redistribution. mechanisms the trigger for the correlation change is the learned stimulus-to-action mapping (the motif signalling which port to peck), not reward pairing, familiarity or exposure, which were equated across motif classes Quotation withheld because this record is not cleared for redistribution. positively correlated noise between similarly tuned neurons impairs coding because pooling cannot remove shared noise without removing signal Quotation withheld because this record is not cleared for redistribution. positively correlated noise between oppositely tuned neurons improves coding because subtraction removes noise and strengthens signal Quotation withheld because this record is not cleared for redistribution. trial shuffling separates the contribution of single-neuron plasticity from that of correlation change, since it leaves individual responses untouched Quotation withheld because this record is not cleared for redistribution. the design equalises reward across all trained motifs so only the stimulus-response association differs Quotation withheld because this record is not cleared for redistribution. testing each motif in isolation confirms the behavioural asymmetry is about which response the motif signals, not about reward Quotation withheld because this record is not cleared for redistribution. The narrow-spiking (putative interneuron) sample was too small for reliable pair analysis, so the population results rest on wide-spiking pairs only. what this limits: every pairwise-correlation result: WE FOCUS ON WS NEURONS ... BECAUSE OUR SAMPLE OF NS NEURONS WAS NOT SUFFICIENT (N = 17 PAIRS FROM 36 SINGLE NEURONS) TO PERFORM RELIABLE ANALYSIS Quotation withheld because this record is not cleared for redistribution. The claim that correlation-structure plasticity is confined to putative excitatory neurons is underpowered. what this limits: the cell-type specificity claim specifically: THIS SUGGESTS THAT THE PLASTICITY ... PRIMARILY EXISTS WITHIN WS (PUTATIVE EXCITATORY) NEURONS, ALTHOUGH MORE DATA ARE NECESSARY - the negative half rests on the underpowered sample Quotation withheld because this record is not cleared for redistribution. Recordings were made under urethane anaesthesia, unlike the awake comparison study, and the circuitry producing the stimulus-specific flexibility is not identified. what this limits: the comparison with the awake primate study: A METHODOLOGICAL DIFFERENCE BETWEEN OUR STUDY AND THAT OF (GU, ET. AL. 2011) IS THAT WE USED ANESTHETIZED ANIMALS WHILE GU AND COLLEAGUES USED AWAKE ANIMALS Quotation withheld because this record is not cleared for redistribution. The local or extrinsic circuit mechanism controlling the correlation structure remains to be established. what this limits: any circuit account of the effect: FURTHER EXPERIMENTS WILL BE NECESSARY TO ELUCIDATE THE CIRCUITRY THAT YIELDS THIS STIMULUS-SPECIFIC FLEXIBILITY Quotation withheld because this record is not cleared for redistribution. narrow-spiking neurons were excluded because the sample was too small for reliable analysis duplicate of: limitations[0] · what this limits: the same scope as limitations[0] - narrow-spiking neurons were excluded because the sample was too small for reliable analysis, so the population results are the broad-spiking ones Quotation withheld because this record is not cleared for redistribution. recordings were made under urethane anaesthesia, unlike the study whose result this contradicts what this limits: the same scope as limitations[2], with the authors' position: WE THINK IT IS HIGHLY UNLIKELY THAT ANESTHESIA COULD ACCOUNT FOR THE DIFFERENCES BETWEEN OUR RESULTS FOR TWO REASONS Quotation withheld because this record is not cleared for redistribution. the single-neuron firing rate difference is modest relative to the behavioural change it accompanies what this limits: any single-neuron reading of the learning effect: THE MODULATION OF SINGLE NEURON FIRING RATES IS SUBTLE ... ESPECIALLY IN LIGHT OF THE ANIMALS' ROBUST CHANGES IN BEHAVIORAL PERFORMANCE Quotation withheld because this record is not cleared for redistribution. the effect in pairs of neurons is described as modest, and only grows large in bigger populations what this limits: the size of the correlation effect at the scale it was measured: A MODEST EFFECT IN PAIRS OF NEURONS GROWS CONSIDERABLY WITH POPULATION SIZE, and for irrelevant motifs the effect IS SMALL AND DOES NOT DEPEND ON POPULATION SIZE Quotation withheld because this record is not cleared for redistribution. the comparison with the contradicting study differs in species, brain region, sensory modality and type of learning what this limits: the disagreement with the primate study, whose differing parameters are enumerated: THIS STUDY DIFFERED FROM OURS IN MULTIPLE WAYS (E.G. SPECIES, BRAIN REGION, AND SENSORY MODALITY) and type of learning Quotation withheld because this record is not cleared for redistribution. the concept of 'occasion setting' is imported to reinterpret the CLM motif representation as a predictive mapping onto the learned response rather than a sensory trace Quotation withheld because this record is not cleared for redistribution. occasion setting - the idea that a stimulus signals which response is appropriate rather than predicting reward itself - used to reinterpret what the cortical representation is for Quotation withheld because this record is not cleared for redistribution. multinomial logistic regression, chosen because stimulus identity is categorical and cannot be parameterised Quotation withheld because this record is not cleared for redistribution. the trial-shuffling repeated 50 times, with the AVERAGE classification probability used rather than a single shuffle eliminates: correlations introduced by the shuffle itself - one shuffle can create spurious structure, so a single shuffled baseline could be beaten or missed by chance · outcome (stated by the tagger, no quotation): held Quotation withheld because this record is not cleared for redistribution. CLM CLM Quotation withheld because this record is not cleared for redistribution. noise correlation is shared trial-to-trial variability; signal correlation is tuning similarity - and the coding consequence of the first depends entirely on the sign of the second Quotation withheld because this record is not cleared for redistribution. task-relevant means the motif signalled which response was correct; task-irrelevant motifs were equally familiar and equally rewarded, so relevance here is purely a stimulus-response association Quotation withheld because this record is not cleared for redistribution. the classification ratio is defined so that above one means correlations help and below one means they hurt, making the sign of the effect readable directly Quotation withheld because this record is not cleared for redistribution. the CLM representation is interpreted as a predictive mapping of the learned response rather than a sensory trace - occasion setting rather than encoding Quotation withheld because this record is not cleared for redistribution. claim_type: mechanism a second reading reaching the same value. Recorded: mechanism — note: a second reading reaching the same value from a different sentence. Not applied - a person decides. Quotation withheld because this record is not cleared for redistribution. method: two-alternative operant training on song motifs, then simultaneous recording of up to 8 neurons and multinomial logistic regression decoding with and without trial shuffling a second reading DIFFERING from the recorded value. Recorded: electrophysiology — note: a second reading, verbatim-verified, disagreeing with the value on the record. Not applied - a person decides. Quotation withheld because this record is not cleared for redistribution. phenomenon: learning-dependent plasticity of inter-neuronal correlation structure a second reading DIFFERING from the recorded value. Recorded: inter-neuronal noise correlations in population coding — note: a second reading, verbatim-verified, disagreeing with the value on the record. Not applied - a person decides. Quotation withheld because this record is not cleared for redistribution. builds-on → Oram et al. (1998) Quotation withheld because this record is not cleared for redistribution. replicates → Cohen and Kohn (2011); Gu et al. (2011); Kohn and Smith (2005); Zohary et al. (1994) Quotation withheld because this record is not cleared for redistribution. replicates → Jeanne et al. (2011); Gentner and Margoliash (2003); Thompson and Gentner (2010) Quotation withheld because this record is not cleared for redistribution. contradicts → the inference from earlier work that stimulus-reward association drives auditory cortical plasticity differs in: THE AXIS IS TASK RELEVANCE OF THE STIMULUS. 'ONLY THE TASK-RELEVANT motifs - those that birds learned to associate with a particular pecking location - elicited neural population activity with a NEGATIVE relationship between signal and noise correlations. In contrast, task-irrelevant motifs ... elicited neural population activity' [with the opposite sign]. The sign of the correlation relationship depends on learned relevance, so a claim about signal-noise correlations is only well-formed once the stimulus class is stated. Quotation withheld because this record is not cleared for redistribution. confirms → the theoretical prediction that changing the sign of the signal-noise correlation relationship could underlie attention or learning Quotation withheld because this record is not cleared for redistribution. qualifies → Gu et al., who found perceptual learning did not alter the slope of this relationship Quotation withheld because this record is not cleared for redistribution. supports → theoretical work showing correlations help when signal and noise correlations have opposite sign Quotation withheld because this record is not cleared for redistribution. noise correlation —contrasted-with→ signal correlation noise correlation signal correlation differs in: THE AXIS DETERMINES THE SIGN OF THE EFFECT, NOT ITS SIZE. 'noise correlations can either enhance or impair coding fidelity, depending on how the noise correlation relates to tuning similarity (KNOWN AS SIGNAL CORRELATION) for each pair of neurons.' A claim that noise correlations help or hurt is not well-formed until this relation is stated - so two papers reporting opposite effects can both be right, and a join on 'noise correlation' alone would make them contradict. · strength of the relation: 1 · note: SIGN: correlated noise IMPAIRS coding between similarly tuned neurons and IMPROVES it between oppositely tuned ones. the same edge has opposite consequences depending on the second variable Quotation withheld because this record is not cleared for redistribution. EVIDENCE CORRECTED (RE-GROUNDED) · task-relevant —contrasted-with→ task-irrelevant task-relevant task-irrelevant EVIDENCE CORRECTED: RE-GROUNDED. The old span said the motifs were heard equally often. This sentence names both AND GIVES THE DIFFERENTIA - 'the primary DIFFERENCE ... was thus the learned association…' between motifs and behavioural choices - so the contrast carries what it consists in. — strength of the relation: 1 · note: matched on familiarity and reward by design, differing only in whether the motif signalled a response - which is what makes the comparison interpretable Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span names only 'task-relevant motifs'; the task-irrelevant motifs are implied by 'only'. Quotation withheld because this record is not cleared for redistribution. noise correlation —causes→ population code noise correlation population code strength of the relation: 0.9 · note: and the effect cannot be predicted from the average responses of individual neurons, which is the paper's motivating point Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the RELATIONSHIP between signal and noise correlations affects the fidelity of representations; causes is loose. Quotation withheld because this record is not cleared for redistribution. population code —operationalised-as→ classification ratio population code classification ratio strength of the relation: 1 Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span writes 'encoding', not 'population code'. Quotation withheld because this record is not cleared for redistribution. task-relevant —contrasted-with→ task-irrelevant task-relevant task-irrelevant strength of the relation: 1 · note: RE-GROUNDS another row of this record: the old span named ONE endpoint. Quotation withheld because this record is not cleared for redistribution. Spearman correlation Spearman correlation Quotation withheld because this record is not cleared for redistribution. Pearson correlation Pearson correlation Quotation withheld because this record is not cleared for redistribution. ANCOVA ANCOVA Quotation withheld because this record is not cleared for redistribution. repeated measures ANOVA repeated measures ANOVA Quotation withheld because this record is not cleared for redistribution. Friedman test Friedman test Quotation withheld because this record is not cleared for redistribution. Kruskal-Wallis test Kruskal-Wallis test Quotation withheld because this record is not cleared for redistribution. t-test t-test Quotation withheld because this record is not cleared for redistribution. Wilcoxon signed-rank test Wilcoxon signed-rank test Quotation withheld because this record is not cleared for redistribution. multinomial logistic regression multinomial logistic regression Quotation withheld because this record is not cleared for redistribution. cross-validation cross-validation Quotation withheld because this record is not cleared for redistribution. trial shuffling (permutation control) trial shuffling (permutation control) Quotation withheld because this record is not cleared for redistribution. Lilliefors normality test Lilliefors normality test Quotation withheld because this record is not cleared for redistribution. Hartigan's dip test for bimodality Hartigan's dip test for bimodality Quotation withheld because this record is not cleared for redistribution. principal components analysis principal components analysis Quotation withheld because this record is not cleared for redistribution. standard error of the mean standard error of the mean Quotation withheld because this record is not cleared for redistribution. multinomial logistic regression decoding multinomial logistic regression decoding Quotation withheld because this record is not cleared for redistribution. trial shuffling trial shuffling Quotation withheld because this record is not cleared for redistribution. two-way ANOVA interaction two-way ANOVA interaction Quotation withheld because this record is not cleared for redistribution. Friedman test Friedman test Quotation withheld because this record is not cleared for redistribution. For motifs that had guided the birds' choices, the usual positive coupling between signal and noise correlation was inverted to a negative one. Quotation withheld because this record is not cleared for redistribution. For task-irrelevant and novel motifs the signal-noise correlation relationship stayed positive. Quotation withheld because this record is not cleared for redistribution. Mean noise correlation did not differ between motif classes. Quotation withheld because this record is not cleared for redistribution. The magnitude of signal correlation was likewise unaffected by task relevance. Quotation withheld because this record is not cleared for redistribution. Task-relevant motifs showed a higher classification ratio than task-irrelevant or novel motifs. Quotation withheld because this record is not cleared for redistribution. The coding benefit conferred by correlations was flat with population size for irrelevant and novel motifs but grew steadily with population size for task-relevant motifs. Quotation withheld because this record is not cleared for redistribution. Firing rates were modestly higher for task-relevant than task-irrelevant motifs, and similar for irrelevant and novel motifs. Quotation withheld because this record is not cleared for redistribution. The authors claim priority for observing a negative signal-noise correlation relationship in cortex. Quotation withheld because this record is not cleared for redistribution. learning produces stimulus-specific changes in inter-neuronal correlations that enhance recognition of behaviourally relevant signals, and the enhancement grows with population size Quotation withheld because this record is not cleared for redistribution. correlations improve encoding when signal and noise correlations have opposite sign and impair it when they share sign Quotation withheld because this record is not cleared for redistribution. task-relevant motifs showed a higher classification ratio than task-irrelevant or novel motifs Quotation withheld because this record is not cleared for redistribution. the benefit of correlations grows with population size for task-relevant motifs and not for the others Quotation withheld because this record is not cleared for redistribution. enhanced coding of task-relevant motifs comes from both single-neuron response properties and correlations between neurons Quotation withheld because this record is not cleared for redistribution. single-neuron firing rate differences were modest relative to the behavioural change Quotation withheld because this record is not cleared for redistribution. task-irrelevant motifs were indistinguishable from novel motifs in correlation structure despite being equally familiar and rewarded Quotation withheld because this record is not cleared for redistribution. a negative relationship between signal and noise correlations had not previously been demonstrated Quotation withheld because this record is not cleared for redistribution. noise correlation noise correlation Quotation withheld because this record is not cleared for redistribution. signal correlation signal correlation Quotation withheld because this record is not cleared for redistribution. classification ratio classification ratio Quotation withheld because this record is not cleared for redistribution. motif motif Quotation withheld because this record is not cleared for redistribution. inverted correlation structure inverted correlation structure Quotation withheld because this record is not cleared for redistribution. wide spiking (WS) and narrow spiking (NS) neurons wide spiking (WS) and narrow spiking (NS) neurons Quotation withheld because this record is not cleared for redistribution. classification ratio classification ratio Quotation withheld because this record is not cleared for redistribution. task-relevant motif task-relevant motif Quotation withheld because this record is not cleared for redistribution. task-irrelevant motif task-irrelevant motif Quotation withheld because this record is not cleared for redistribution. behavioral relevance behavioral relevance Quotation withheld because this record is not cleared for redistribution. number of neurons number of neurons Quotation withheld because this record is not cleared for redistribution. noise correlation noise correlation Quotation withheld because this record is not cleared for redistribution. signal correlation signal correlation Quotation withheld because this record is not cleared for redistribution. classification ratio classification ratio Quotation withheld because this record is not cleared for redistribution. motif assignment motif assignment Quotation withheld because this record is not cleared for redistribution. analysis window analysis window Quotation withheld because this record is not cleared for redistribution. task relevance of the motif task relevance of the motif Quotation withheld because this record is not cleared for redistribution. population size population size Quotation withheld because this record is not cleared for redistribution. r = −0.15, p = 0.051 — correlation between signal correlation and noise correlation across neuron pairs for task-relevant motifs (negative) Quotation withheld because this record is not cleared for redistribution. r = 0.19 and r = 0.23 — the same signal-noise coupling for task-irrelevant and for novel motifs (positive) Quotation withheld because this record is not cleared for redistribution. p = 7.9×10 −5 — significance of the difference in regression slopes between motif classes Quotation withheld because this record is not cleared for redistribution. 0.082 / 0.100 / 0.087 — mean noise correlation for task-relevant, task-irrelevant and novel motifs - small and positive in every class Quotation withheld because this record is not cleared for redistribution. n = 176 pairs from 98 single neurons — the analysed sample of wide-spiking neuron pairs (narrow-spiking sample of 17 pairs was too small to analyse) Quotation withheld because this record is not cleared for redistribution. 51% vs 25% — decoding accuracy of the example two-neuron classifier against chance for four motifs Quotation withheld because this record is not cleared for redistribution. 5–6 neurons — population size at which the coding benefit of the learned correlation structure becomes substantial Quotation withheld because this record is not cleared for redistribution. Nine — number of wild-caught European starlings trained on the two-alternative choice task Quotation withheld because this record is not cleared for redistribution. 176 pairs from 98 wide-spiking neurons, against 17 pairs from 36 narrow-spiking — the analysed sample and the excluded one Quotation withheld because this record is not cleared for redistribution. EVIDENCE CORRECTED (RE-GROUNDED - THE POINTER WAS WRONG, NOT THE CLAIM) · 3.03 versus 2.74 Hz, p = 0.0080 — the modest single-neuron firing rate difference EVIDENCE CORRECTED: RE-GROUNDED - THE POINTER WAS WRONG, NOT THE CLAIM. The old evidence was a hedging remark - 'The modulation of single neuron firing rates is subtle' - carrying none of 3.03, 2.74 or 0.0080. The sentence that states the comparison is now the evidence. Quotation withheld because this record is not cleared for redistribution. p = 1.8e-8 for the interaction — the dependence of the coding effect on the sign relationship Quotation withheld because this record is not cleared for redistribution. sets of up to 8 neurons — the population sizes available for the scaling analysis Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.1016/j.neuron.2013.02.023.