KumarpersName, BonnicipersName, TekipersName, AguspersName, PressnitzerpersName, MaguirepersName, GriffithspersName (2014date). “Representations of specific acoustic patterns in the auditory cortex and hippocampus.” Proceedings of the Royal Society B: Biological Sciencestitle·j.
Object of study — where in the brain a freshly and implicitly learnt noise-like acoustic patternterm is represented
Research question — how the brain learns novel acoustic patternsterm, and whether a mechanism exists for storing acoustic patternsterm as such
Contribution — shows that planum temporaleterm and hippocampusterm, but not primary auditory cortex, carry activity patterns distinguishing three specific learnt noise exemplars
phenomenon · claim-type · cross-domain-transfer · mechanism · claim-relation · replication-status · limitation — grounded in the full text| phenomenon | implicit learning of random acoustic patterns — “Previous behavioural studies have shown that repeated presentation of a randomly chosen acoustic pattern leads to the unsupervised learning of some of its specific acoustic features.” | |
| claim-type | existence — “Our results demonstrate that these structures are involved in the neural representation of specific acoustic patterns after they have been learnt.” | |
| method — evidence | — “We used high-resolution functional magnetic resonance imaging (fMRI) and multi-voxel pattern analysis (MVPA).” | |
| object of study — evidence | — “The objective of our study was to determine the neural substrate for the representation of freshly learnt acoustic patterns.” | |
| research question — evidence | — “Although a number of brain imaging studies [ 1 – 4 ] have highlighted the brain system which represents the meaning of sounds, key questions remain as to how the brain learns novel acoustic patterns, and whether a specific mechanism exists for the storage of acoustic patterns per se .” | |
| contribution — evidence | — “We found that activity in planum temporale and the hippocampus reliably distinguished between the learnt acoustic patterns. Our results demonstrate that these structures are involved in the neural representation of specific acoustic patterns after they have been learnt.” | |
| transfer | from machine learning: a linear support vector classifier with cross-validation and searchlight feature selection, used to ask whether a region distinguishes stimuli rather than whether it activates more — “The purpose of feature selection is to reduce the set of features (in this case, voxels) in a dataset to those most likely to carry relevant information. This is effectively the same as removing voxels most likely to carry noise and is a way of increasing the signal-to-noise ratio.” | transfers[0] |
| transfer | from psychoacoustics: the repeated-noise paradigm, in which a silently recurring random exemplar reveals template learning through improved repetition detection — “However, without this being mentioned to subjects, there was in fact a third type of trials: one particular exemplar of repeated noise was presented over several trials, randomly interspersed throughout an experimental block. These trials (reference repeated noise, RefRN) were thus initially drawn from the same process as RN but, unlike RN, they were acoustically identical across several trials.” | transfers[1] |
| mechanism | A connected planum temporale-hippocampus system converts the dense HG activity pattern for a complex sound into a sparser representation that can be stored in long-term memory — “The purpose of such a system may be to transform the rich pattern of activity for complex sounds expressed in HG into sparser representations, more amenable to long-term memory storage.” | mechanisms[0] |
| mechanism | Repeated exposure lays down a separate memory trace for each learnt exemplar in planum temporale and hippocampus, which is what the classifier reads out — “The most parsimonious explanation for our data is a unique memory trace in PT and HC that is formed for each of the three RefTCs, although more work will be needed to probe this point further.” | mechanisms[1] |
| mechanism | the learning is made incidental by construction: one exemplar silently recurs across trials while every other stimulus is generated anew, so exposure is the only difference — “Without the subjects’ knowledge, however, one exemplar of a repeated tone cloud (reference tone cloud, RefTC) remained the same from trial to trial. That is, for the RefTC stimulus not only the same burst of tone cloud was repeated (three times) within a trial, but also the exact same exemplar reoccurred over several trials. The trials consisting of RefTC stimuli were presented randomly among trials of RTC and NTC. A RefTC stimulus was drawn from the same statistical process as any other RTC; the only difference was that the same exemplar of RefTC was presented across different trials.” | mechanisms[2] |
| mechanism | the repeated and non-repeated stimuli are drawn from the same statistical process, so acoustic parameters cannot explain the difference — “Importantly, the repeatedly presented stimuli (RefTCs) and the non-repeated stimuli (RTC) were well balanced with respect to acoustic parameters. The key difference between the two conditions was in terms of exposure over the course of the experiment.” | mechanisms[3] |
| mechanism | training is run inside the scanner while it is running, so the acoustic environment of learning matches that of testing — “All training occurred inside the MRI scanner while it was running in order to create the same conditions during learning as during subsequent testing.” | mechanisms[4] |
| mechanism | the analysis was switched from univariate to multivariate after the first found nothing, and the switch is disclosed — “We initially used standard univariate analyses to interrogate the data, but did not find any significant difference in the mean activity evoked by the three RefTC stimuli in any part of the brain. We therefore focused on using MVPA which we believed would have increased sensitivity in our experimental context.” | mechanisms[5] |
| mechanism | a boundary with no consensus in the literature is chosen explicitly rather than left implicit — “The markers for the posterior boundary of PT are not well established, because in most cases the posterior end of the sylvian fissure bifurcates into ascending and descending rami, the pattern of which varies across subjects [ 21 ]. There is no consensus on whether the posterior end of PT is limited to the posterior end of the horizontal portion of sylvian fissure [ 22 , 23 ] or extends all the way up to the end-point of the ascending ramus (which could cover a part of the parietal lobe [ 21 ]). In this study, we chose the latter option.” | mechanisms[6] |
| mechanism | label shuffling is used to confirm the classifier is not exploiting some artefact of the design — “To ensure that the classifiers were unbiased and that classification was based on unique representations for each of the three RefTCs, we ran a control analysis in which we randomly shuffled (from trial to trial) the labels of the three RefTCs. The classifiers were trained and tested as above. As expected, the performance of the classifier dropped to chance level for all the brain regions” | mechanisms[7] |
| mechanism | a second control classifies stimuli whose exemplar changes every trial, which should be undecodable if the effect really depends on repetition — “To further confirm that classification of RefTCs was based on stable representations (owing to the repeated presentation of the same exemplars), we ran a second control analysis. In this analysis, we divided the 60 RTC trials randomly into three classes and performed the same classification analysis. Since a different exemplar is presented in every trial of RTC, representation of RTC changes from trial to trial. It is therefore expected that the classifier should not be able to classify the RTC stimuli significantly better than chance.” | mechanisms[8] |
| mechanism | regions of interest are hand-segmented per subject from that subject's own anatomy rather than warped to a template — “The markers for the posterior boundary of PT are not well established, because in most cases the posterior end of the sylvian fissure bifurcates into ascending and descending rami, the pattern of which varies across subjects [ 21 ]. There is no consensus on whether the posterior end of PT is limited to the posterior end of the horizontal portion of sylvian fissure [ 22 , 23 ] or extends all the way up to the end-point of the ascending ramus (which could cover a part of the parietal lobe [ 21 ]). In this study, we chose the latter option.” | mechanisms[9] |
| mechanism | feature selection by searchlight is used to raise signal-to-noise before classification, fitted on the training split only — “The purpose of feature selection is to reduce the set of features (in this case, voxels) in a dataset to those most likely to carry relevant information. This is effectively the same as removing voxels most likely to carry noise and is a way of increasing the signal-to-noise ratio.” | mechanisms[10] |
| mechanism | the sensory and timbre explanations are argued against by absence of precedent rather than by measurement, and the authors say so — “We considered whether our results might be explained by sensory or perceptual representations in PT and/or HC. The existence of sensory representations of fine spectro-temporal features that occurs in non-primary auditory cortex and HC but not in primary auditory cortex is unlikely. Mapping of the different perceptual timbre of different exemplars could occur in non-primary auditory cortex, but there is no precedent for any effect of manipulating spectro-temporal structure and timbre on hippocampal activity” | mechanisms[11] |
| mechanism | the regions are predicted in advance with reasons, so the six-way comparison is hypothesis-driven rather than exploratory — “We also hypothesized that structures in the medial temporal lobes (MTLs) would be crucial. In addition to hippocampus (HC), which has been shown to have a role in long-term memory for sounds [ 15 , 16 ], we also speculated that entorhinal/perirhinal cortex (EPC) and parahippocampal cortex (PHC), both of which receive dense input from the non-primary auditory areas [ 17 ], may be involved.” | mechanisms[12] |
| replicates | Agus et al. [5] — “As in Agus et al. [ 5 ], the behavioural data in our study revealed better performance on repeated exemplars compared to performance on non-repeated exemplars confirming learning of acoustic patterns.” | claim_relations[0] |
| extends | the magnetoencephalography study [37] showing phase-pattern specificity for learnt noise exemplars — “However, compared to the results in [ 37 ], our results are more specific with respect to the areas of auditory cortex involved in storing pre-semantic-specific templates.” | claim_relations[1] |
| extends | visual-domain studies [39] showing hippocampal sensitivity to low-level perceptual features — “The results of our study further extend these results by showing that the HC is not only sensitive to low-level features of stimuli that are explicitly recalled, but also constructs representations that are specific to acoustic features which are learned implicitly.” | claim_relations[2] |
| extends | behavioural work showing that repeated presentation of a random noise exemplar produces unsupervised template learning — “The main result was that performance was considerably better for the RefRN stimuli than for the RN stimuli. Since the only difference between RefRN and RN was that RefRN was heard over several trials, the improved performance could be attributed to the learning of a template for RefRN.” | claim_relations[3] |
| extends | a magnetoencephalography study showing different phase patterns for different learnt noise exemplars, by localising the representation more specifically — “Although HC is known to be involved in processing of complex and meaningful sounds [ 15 , 38 ], to the best of our knowledge, our study is the first to demonstrate the encoding of noise-like acoustic patterns in the HC.” | claim_relations[4] |
| extends | visual work showing hippocampal sensitivity to low-level perceptual features, to implicitly learnt acoustic features — “The results of our study further extend these results by showing that the HC is not only sensitive to low-level features of stimuli that are explicitly recalled, but also constructs representations that are specific to acoustic features which are learned implicitly.” | claim_relations[5] |
| builds-on | the hypothesis that planum temporale generates acoustic templates at a stage before semantic processing — “Specifically, we predicted engagement of planum temporale (PT) which has been hypothesized to be involved in the generation of acoustic ‘templates’ at a stage before semantic processing [ 12 ] and superior temporal sulcus (STS), which has been shown to store long-term memories for sounds” | claim_relations[6] |
| builds-on | monkey tracing showing belt and parabelt, but not core, project directly to entorhinal cortex — “However, the connectivity pattern between auditory cortex and the MTL of monkeys [ 47 ] (for review, see [ 17 ]) shows that belt and parabelt, but not the core, of the auditory cortex have direct projections to the entorhinal cortex.” | claim_relations[7] |
| supports | evidence that primary auditory cortex stores long-term representations of specific auditory experiences — “Given evidence for the involvement of primary auditory cortex in storing long-term representations of specific auditory experiences [ 9 – 11 ], we hypothesized that Heschl's gyrus (HG), which contains primary auditory cortex, would be involved.” | claim_relations[8] |
| sense | planum temporale - bounded at the horizontal sylvian fissure versus extending to the ascending ramus and into parietal lobe; no consensus exists and the choice is stipulated per study — “The markers for the posterior boundary of PT are not well established, because in most cases the posterior end of the sylvian fissure bifurcates into ascending and descending rami, the pattern of which varies across subjects [ 21 ]. There is no consensus on whether the posterior end of PT is limited to the posterior end of the horizontal portion of sylvian fissure [ 22 , 23 ] or extends all the way up to the end-point of the ascending ramus (which could cover a part of the parietal lobe [ 21 ]). In this study, we chose the latter option.” | senses[0] |
| sense | learning - explicit and reportable versus implicit; probed here with one debriefing question, and the conclusion hedged to some cases — “Subjective reports from the participants after the experiment showed that most were not aware of any stimulus re-occurring during the experiment, suggesting that in some cases the learning occurred implicitly.” | senses[1] |
| sense | the same acoustic structure - as a repeated stimulus within a trial versus as the same exemplar recurring across trials; the second is what produces learning and the first is the control — “Without the subjects’ knowledge, however, one exemplar of a repeated tone cloud (reference tone cloud, RefTC) remained the same from trial to trial. That is, for the RefTC stimulus not only the same burst of tone cloud was repeated (three times) within a trial, but also the exact same exemplar reoccurred over several trials. The trials consisting of RefTC stimuli were presented randomly among trials of RTC and NTC. A RefTC stimulus was drawn from the same statistical process as any other RTC; the only difference was that the same exemplar of RefTC was presented across different trials.” | senses[2] |
| sense | an analysis result - as absent under mean activation versus present under distributed pattern classification, on the same scans — “We initially used standard univariate analyses to interrogate the data, but did not find any significant difference in the mean activity evoked by the three RefTC stimuli in any part of the brain. We therefore focused on using MVPA which we believed would have increased sensitivity in our experimental context.” | senses[3] |
| sense | a representation of a sound - as a sensory or timbre code versus as a memory trace; the second is preferred on parsimony rather than measurement — “The most parsimonious explanation for our data is a unique memory trace in PT and HC that is formed for each of the three RefTCs, although more work will be needed to probe this point further.” | senses[4] |
| control | a control analysis with the class labels RANDOMLY SHUFFLED from trial to trial, run through the same classifier — held — rules out classifier bias and any structure the pipeline itself imposes - a decoder given shuffled labels should perform at chance, and anything above chance is the method rather than the representation — “To ensure that the classifiers were unbiased and that classification was based on unique representations for each of the three RefTCs, we ran a control analysis in which we randomly shuffled” | controls[0] |
| edge | reference tone cloud —contrasted-with→ repeated tone cloud — “Without the subjects’ knowledge, however, one exemplar of a repeated tone cloud (reference tone cloud, RefTC) remained the same from trial to trial. That is, for the RefTC stimulus not only the same burst of tone cloud was repeated (three times) within a trial, but also the exact same exemplar reoccurred over several trials. The trials consisting of RefTC stimuli were presented randomly among trials of RTC and NTC. A RefTC stimulus was drawn from the same statistical process as any other RTC; the only difference was that the same exemplar of RefTC was presented across different trials.” | concept_edges[0] |
| audit | planum temporale —is-a→ computational hub — Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the findings are 'CONSISTENT WITH a role as' a computational hub; is-a is hedged. — “The availability of pattern specific information in the PT is consistent with a role as a computational hub [ 12 ], where spectro-temporal patterns are matched with learned patterns that are stored beyond the auditory cortex.” | concept_edges[1] |
| weight | planum temporale —is-a→ computational hub — 0.8 — “The availability of pattern specific information in the PT is consistent with a role as a computational hub [ 12 ], where spectro-temporal patterns are matched with learned patterns that are stored beyond the auditory cortex.” | concept_edges[1] |
| weight | planum temporale —contrasted-with→ Heschl's gyrus — 0.95 — “We found that patterns of activity across voxels in PT, but not in HG, could distinguish between the three learned acoustic patterns. We speculate that this might reflect a type of representation within PT that is not a simple representation of acoustic pattern but a more refined representation that requires interaction with HC (see below).” | concept_edges[2] |
| audit | multi-voxel pattern analysis —operationalised-as→ support vector machine — Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but endpoint by anaphora: the section is headed 'Multi-voxel pattern analysis' (previous line). — “A linear support vector machine (SVM) classifier was created for each ROI.” | concept_edges[3] |
| weight | multi-voxel pattern analysis —operationalised-as→ support vector machine — 0.95 — “A linear support vector machine (SVM) classifier was created for each ROI.” | concept_edges[3] |
| audit | superior temporal sulcus —causes→ categorical perception — Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the STS 'IS INVOLVED IN' categorical perception; causes is loose. — “From the auditory perception point of view, converging evidence from neuroimaging studies show that STS is involved in categorical perception of speech [ 13 ] and non-speech [ 14 , 41 ] auditory stimuli.” | concept_edges[4] |
| weight | superior temporal sulcus —causes→ categorical perception — 0.7 — “From the auditory perception point of view, converging evidence from neuroimaging studies show that STS is involved in categorical perception of speech [ 13 ] and non-speech [ 14 , 41 ] auditory stimuli.” | concept_edges[4] |
| audit | hippocampus —causes→ memory trace — Audit 2026-09-28: DOUBTFUL - read against its span and kept, but the span offers a memory trace as the 'most parsimonious explanation', needing more work; causes is hedged. — “The most parsimonious explanation for our data is a unique memory trace in PT and HC that is formed for each of the three RefTCs, although more work will be needed to probe this point further.” | concept_edges[5] |
| note | hippocampus —causes→ memory trace — DOUBTFUL, found by an audit that read the edge against its span: 'formed in' the hippocampus - locus, not cause. The edge is kept live; a reader should weigh it with this in mind. — “The most parsimonious explanation for our data is a unique memory trace in PT and HC that is formed for each of the three RefTCs, although more work will be needed to probe this point further.” | concept_edges[5] |
| weight | hippocampus —causes→ memory trace — 0.8 — “The most parsimonious explanation for our data is a unique memory trace in PT and HC that is formed for each of the three RefTCs, although more work will be needed to probe this point further.” | concept_edges[5] |
| outcome (stated by the tagger, no quotation) | control — held | controls[0] |
| for which use | limitation — the negative STS result only: THIS LACK OF SIGNIFICANCE MAY BE OWING TO THE LOW POWER OF OUR STUDY, SO A ROLE FOR STS ... CANNOT BE RULED OUT — “This lack of significance may be owing to the low power of our study, so a role for STS in the learning of novel acoustic patterns cannot be ruled out.” | limitations[0] |
| for which use | limitation — the memory-trace interpretation, not the decoding result: THE MOST PARSIMONIOUS EXPLANATION ... ALTHOUGH MORE WORK WILL BE NEEDED TO PROBE THIS — “The most parsimonious explanation for our data is a unique memory trace in PT and HC that is formed for each of the three RefTCs, although more work will be needed to probe this point further.” | limitations[1] |
| for which use | limitation — any claim about when or how the representation forms: THE FMRI DATA ACQUIRED DURING THIS PHASE WERE NOT INTENDED (NOR WERE THEY SUITABLE) FOR multivariate analysis — “Our learning phase was optimized to facilitate acquisition of the three RefTCs; consequently, the fMRI data acquired during this phase were not intended (nor were they suitable) for analysis using MVPA, because our focus was solely on where the learnt RefTCs were represented.” | limitations[2] |
| for which use | limitation — the implicit-learning claim specifically: awareness was probed with a single post-hoc question and two of seven listeners reported some awareness — “Subjective reports from the participants after the experiment showed that most were not aware of any stimulus re-occurring during the experiment, suggesting that in some cases the learning occurred implicitly.” | limitations[3] |
| for which use | limitation — the same scope as limitations[0] - the near-significant STS result is attributed to low power, which leaves a role there unexcluded rather than excluded — “This lack of significance may be owing to the low power of our study, so a role for STS in the learning of novel acoustic patterns cannot be ruled out.” | limitations[4] |
| duplicate of | limitation — limitations[2] — “Our learning phase was optimized to facilitate acquisition of the three RefTCs; consequently, the fMRI data acquired during this phase were not intended (nor were they suitable) for analysis using MVPA, because our focus was solely on where the learnt RefTCs were represented.” | limitations[5] |
| for which use | limitation — the same scope as limitations[2] - the learning-phase scans were not suitable for the main analysis, so the time course of learning could not be examined — “Our learning phase was optimized to facilitate acquisition of the three RefTCs; consequently, the fMRI data acquired during this phase were not intended (nor were they suitable) for analysis using MVPA, because our focus was solely on where the learnt RefTCs were represented.” | limitations[5] |
| for which use | limitation — the training protocol behind every stable-representation result: FURTHER WORK IS REQUIRED TO DETERMINE HOW MANY REPETITIONS OF EACH EXEMPLAR ARE REQUIRED DURING THE TRAINING PHASE — “Further work is required to determine how many repetitions of each exemplar are required during the training phase before a stable representation is built in the PT and HC.” | limitations[6] |
| for which use | limitation — the PT-HC relationship: A FURTHER QUESTION CONCERNS THE DIRECTION OF CAUSAL INFLUENCES OF THE PT AND HC ON EACH OTHER — “A further question concerns the direction of causal influences of the PT and HC on each other that occur during the construction of stable representations.” | limitations[7] |
| for which use | limitation — the anatomical argument, which rests on another species: STRUCTURAL CONNECTIVITY BETWEEN HUMAN AUDITORY CORTEX AND THE MTLs IS NOT COMPLETELY UNDERSTOOD — “Structural connectivity between human auditory cortex and the MTLs is not completely understood.” | limitations[8] |
| for which use | limitation — the same scope as limitations[1] - the memory-trace interpretation is offered as the most parsimonious account rather than as an established one — “The most parsimonious explanation for our data is a unique memory trace in PT and HC that is formed for each of the three RefTCs, although more work will be needed to probe this point further.” | limitations[9] |
| replication | untested | |
| limitation | The near-significant STS classification may reflect low statistical power, so an STS role cannot be excluded. — “This lack of significance may be owing to the low power of our study, so a role for STS in the learning of novel acoustic patterns cannot be ruled out.” | limitations[0] |
| limitation | The memory-trace interpretation of the PT and HC decoding is the most parsimonious account but is not established by these data alone. — “The most parsimonious explanation for our data is a unique memory trace in PT and HC that is formed for each of the three RefTCs, although more work will be needed to probe this point further.” | limitations[1] |
| limitation | The learning-phase fMRI data were not suitable for MVPA, so the time course of representation formation could not be examined. — “Our learning phase was optimized to facilitate acquisition of the three RefTCs; consequently, the fMRI data acquired during this phase were not intended (nor were they suitable) for analysis using MVPA, because our focus was solely on where the learnt RefTCs were represented.” | limitations[2] |
| limitation | awareness was probed with a single post-hoc question and two of seven listeners reported some, so the implicitness claim is hedged to some cases — “Subjective reports from the participants after the experiment showed that most were not aware of any stimulus re-occurring during the experiment, suggesting that in some cases the learning occurred implicitly.” | limitations[3] |
| limitation | the near-significant superior temporal sulcus result is attributed to low power, leaving a role there unexcluded — “This lack of significance may be owing to the low power of our study, so a role for STS in the learning of novel acoustic patterns cannot be ruled out.” | limitations[4] |
| limitation | the learning-phase scans were not suitable for the main analysis, so the time course of learning could not be examined — “Our learning phase was optimized to facilitate acquisition of the three RefTCs; consequently, the fMRI data acquired during this phase were not intended (nor were they suitable) for analysis using MVPA, because our focus was solely on where the learnt RefTCs were represented.” | limitations[5] |
| limitation | how many repetitions are needed to build a stable representation is unknown — “Further work is required to determine how many repetitions of each exemplar are required during the training phase before a stable representation is built in the PT and HC.” | limitations[6] |
| limitation | the direction of causal influence between the two regions is unresolved — “A further question concerns the direction of causal influences of the PT and HC on each other that occur during the construction of stable representations.” | limitations[7] |
| limitation | human auditory-to-medial-temporal connectivity is not fully known, so the anatomical argument rests on monkey work — “Structural connectivity between human auditory cortex and the MTLs is not completely understood.” | limitations[8] |
| limitation | the memory-trace interpretation is offered as the most parsimonious rather than as established — “The most parsimonious explanation for our data is a unique memory trace in PT and HC that is formed for each of the three RefTCs, although more work will be needed to probe this point further.” | limitations[9] |
finding · quantity · variable · definition · key-term — grounded in the full text| finding | Of six regions of interest, only the hippocampus and planum temporale allowed a classifier to tell apart which of the three learnt tone-cloud exemplars was being heard; HG, entorhinal/perirhinal cortex and parahippocampal cortex did not. — “Only two regions, HC and PT, showed performance above chance: (HC: t 6 = 2.711, p = 0.018; EPC: t 6 = 0.188, p = 0.43; PHC: t 6 = 0.117, p = 0.46; HG: t 6 = 0.620, p = 0.28; PT: t 6 = 5.106, p = 0.001; STS: t 6 = 1.715, p = 0.07).” | content_tags.findings[0] |
| finding | Subjects detected repetition better for all three repeatedly presented reference tone clouds than for freshly generated repeated tone clouds, showing that the specific patterns had been learnt. — “Post hoc comparison revealed better performance on all the RefTC stimuli compared to RTC (RefTC-1 > RTC, t 6 = 3.96, p = 0.007; RefTC-2 > RTC, t 6 = 4.30, p = 0.005; RefTC-3 > RTC, t 6 = 5.43, p = 0.001).” | content_tags.findings[1] |
| finding | Standard univariate analysis found no difference in mean activity between the three learnt patterns anywhere in the brain, so the effect existed only in the multi-voxel activity pattern. — “We initially used standard univariate analyses to interrogate the data, but did not find any significant difference in the mean activity evoked by the three RefTC stimuli in any part of the brain.” | content_tags.findings[2] |
| finding | When the 60 trial-unique repeated tone clouds were split into three arbitrary classes, classification fell to chance in every region, confirming that decoding depended on stable repeated exemplars rather than on stimulus class labels. — “As predicted, the classifier performance was at chance for all ROIs, (HC: t 6 = −0.058, p = 0.955; EPC: t 6 = −0.421, p = 0.689; PHG: t 6 = 0.953, p = 0.377; HG: t 6 = 1.365, p = 0.221; PT: t 6 = −0.174, p = 0.868; STS: t 6 = −0.316, p = 0.763).” | content_tags.findings[3] |
| finding | Most participants were unaware that any stimulus recurred, suggesting that the learning was sometimes implicit. — “Subjective reports from the participants after the experiment showed that most were not aware of any stimulus re-occurring during the experiment, suggesting that in some cases the learning occurred implicitly.” | content_tags.findings[4] |
| finding | The authors claim this is the first demonstration that the hippocampus encodes noise-like acoustic patterns that carry no semantic content. — “Although HC is known to be involved in processing of complex and meaningful sounds [ 15 , 38 ], to the best of our knowledge, our study is the first to demonstrate the encoding of noise-like acoustic patterns in the HC.” | content_tags.findings[5] |
| finding | only hippocampus and planum temporale classified the three learnt exemplars above chance — “Only two regions, HC and PT, showed performance above chance: (HC: t 6 = 2.711, p = 0.018; EPC: t 6 = 0.188, p = 0.43; PHC: t 6 = 0.117, p = 0.46; HG: t 6 = 0.620, p = 0.28; PT: t 6 = 5.106, p = 0.001; STS: t 6 = 1.715, p = 0.07).” | content_tags.findings[6] |
| finding | planum temporale carried more information than HG, entorhinal/perirhinal and parahippocampal cortex, and hippocampus more than HG — “A repeated measures ANOVA showed a significant effect for region ( F 5,30 = 3.271, p = 0.018), which was driven by more information being present in the HC when compared with HG (HC > HG: t 6 = 2.765, p = 0.033) as well as more information present in PT when compared with HG, EPC and PHC (PT > HG: t 6 = 4.182, p = 0.006; PT > EPC: t 6 = 3.059, p = 0.022; PT > PHC: t 6 = 3.291, p = 0.017).” | content_tags.findings[7] |
| finding | primary auditory cortex did not distinguish the learnt patterns — “We found that patterns of activity across voxels in PT, but not in HG, could distinguish between the three learned acoustic patterns. We speculate that this might reflect a type of representation within PT that is not a simple representation of acoustic pattern but a more refined representation that requires interaction with HC (see below).” | content_tags.findings[8] |
| finding | superior temporal sulcus fell just short of significance at p = 0.07 — “Only two regions, HC and PT, showed performance above chance: (HC: t 6 = 2.711, p = 0.018; EPC: t 6 = 0.188, p = 0.43; PHC: t 6 = 0.117, p = 0.46; HG: t 6 = 0.620, p = 0.28; PT: t 6 = 5.106, p = 0.001; STS: t 6 = 1.715, p = 0.07).” | content_tags.findings[9] |
| finding | stimulus type had a significant effect on detection d-prime — “Analysis of behavioural performance during testing showed a significant effect of stimulus on dprimes ( F 3,18 = 8.98, p = 0.001).” | content_tags.findings[10] |
| finding | shuffling the exemplar labels dropped classification to chance in every region — “To ensure that the classifiers were unbiased and that classification was based on unique representations for each of the three RefTCs, we ran a control analysis in which we randomly shuffled (from trial to trial) the labels of the three RefTCs. The classifiers were trained and tested as above. As expected, the performance of the classifier dropped to chance level for all the brain regions” | content_tags.findings[11] |
| finding | because the RTC exemplar changes every trial, classifying RTC trials split into three classes was expected to be at chance — “To further confirm that classification of RefTCs was based on stable representations (owing to the repeated presentation of the same exemplars), we ran a second control analysis. In this analysis, we divided the 60 RTC trials randomly into three classes and performed the same classification analysis. Since a different exemplar is presented in every trial of RTC, representation of RTC changes from trial to trial. It is therefore expected that the classifier should not be able to classify the RTC stimuli significantly better than chance.” | content_tags.findings[12] |
| finding | five of seven listeners reported no awareness that any sound recurred; one said yes and one said some — “After the MRI scanning, listeners were debriefed by means of a questionnaire. In particular, a question was asked as to whether they thought that any of the sounds recurred during the experiment. One listener said yes, one said some and the remaining five said no.” | content_tags.findings[13] |
| finding | univariate analysis found no difference in mean activity anywhere in the brain — “We initially used standard univariate analyses to interrogate the data, but did not find any significant difference in the mean activity evoked by the three RefTC stimuli in any part of the brain. We therefore focused on using MVPA which we believed would have increased sensitivity in our experimental context.” | content_tags.findings[14] |
| finding | this is reported as the first demonstration of noise-like acoustic patterns encoded in the hippocampus — “Although HC is known to be involved in processing of complex and meaningful sounds [ 15 , 38 ], to the best of our knowledge, our study is the first to demonstrate the encoding of noise-like acoustic patterns in the HC.” | content_tags.findings[16] |
| quantity | 33% — chance classification level, because the classifier discriminated three learnt exemplars — “The classifier accuracy values for each brain region were compared to chance, which in this case was 33% as we were classifying between three exemplars.” | content_tags.quantities[0] |
| quantity | Seven healthy subjects — sample size of the fMRI study — “Seven healthy subjects (two females, mean age = 22.85 years, s.d. = 1.67 years, range = 20–24 years) with no prior history of neurological and psychiatric disorders participated in the study.” | content_tags.quantities[1] |
| quantity | one said yes, one said some and the remaining five said no — debriefing responses on whether subjects noticed any sound recurring, i.e. awareness of the repeated exemplars — “One listener said yes, one said some and the remaining five said no.” | content_tags.quantities[2] |
| quantity | 20 trials each ... 60 trials of RTC and 120 trials of NTC — trial counts per stimulus type in the scanned test session — “After training in the MRI scanner, subjects were tested in a single session consisting of 20 trials each for the three trained RefTC exemplars randomly presented with 60 trials of RTC and 120 trials of NTC (which were generated anew).” | content_tags.quantities[3] |
| quantity | PT 3967.07 (719.66) — mean volume of the planum temporale region of interest in cubic millimetres, with its standard deviation — “Mean volumes (in cubic millimetre, summed across both hemispheres) and standard deviations (SD) for the region of interests (ROIs) were as follows: HC 4188.8 (472.97), EPC 5026.04 (488.71), PHC 1799.29 (256.69), HG 2317.93 (521.93), PT 3967.07 (719.66) and STS 14396.07 (2476.29).” | content_tags.quantities[4] |
| quantity | seven subjects — the sample, mean age 22.85 — “Seven healthy subjects (two females, mean age = 22.85 years, s.d. = 1.67 years, range = 20–24 years) with no prior history of neurological and psychiatric disorders participated in the study.” | content_tags.quantities[5] |
| quantity | 33% — chance classification accuracy for three exemplars — “The classifier accuracy values for each brain region were compared to chance, which in this case was 33% as we were classifying between three exemplars. Given that we were interested in whether results were significantly above chance, one tailed t -tests were used.” | content_tags.quantities[6] |
| quantity | one yes, one some, five no — post-scan awareness responses across the seven listeners — “After the MRI scanning, listeners were debriefed by means of a questionnaire. In particular, a question was asked as to whether they thought that any of the sounds recurred during the experiment. One listener said yes, one said some and the remaining five said no.” | content_tags.quantities[7] |
| quantity | p = 0.07 — superior temporal sulcus classification, just short of the threshold — “Only two regions, HC and PT, showed performance above chance: (HC: t 6 = 2.711, p = 0.018; EPC: t 6 = 0.188, p = 0.43; PHC: t 6 = 0.117, p = 0.46; HG: t 6 = 0.620, p = 0.28; PT: t 6 = 5.106, p = 0.001; STS: t 6 = 1.715, p = 0.07).” | content_tags.quantities[8] |
| variable | dprime (dependent) — “Analysis of behavioural performance during testing showed a significant effect of stimulus on dprimes ( F 3,18 = 8.98, p = 0.001).” | content_tags.variables[2] |
| variable | acoustic parameters (control) — “Importantly, the repeatedly presented stimuli (RefTCs) and the non-repeated stimuli (RTC) were well balanced with respect to acoustic parameters.” | content_tags.variables[3] |
| variable | movement parameters (covariate) — “Participant-specific movement parameters were included as regressors of no interest.” | content_tags.variables[4] |
| variable | classification accuracy — “Classification accuracy for the six ROIs is shown in figure 4 . Only two regions, HC and PT, showed performance above chance: (HC: t 6 = 2.711, p = 0.018; EPC: t 6 = 0.188, p = 0.43; PHC: t 6 = 0.117, p = 0.46; HG: t 6 = 0.620, p = 0.28; PT: t 6 = 5.106, p = 0.001; STS: t 6 = 1.715, p = 0.07).” | content_tags.variables[5] |
| variable | region of interest — “The classifier accuracy values for each brain region were compared to chance, which in this case was 33% as we were classifying between three exemplars. Given that we were interested in whether results were significantly above chance, one tailed t -tests were used.” | content_tags.variables[6] |
| variable | dprimes — “Analysis of behavioural performance during testing showed a significant effect of stimulus on dprimes ( F 3,18 = 8.98, p = 0.001).” | content_tags.variables[7] |
| definition | tone clouds — “The stimuli were ‘tone clouds’. These are noise-like stimuli with a coarser spectro-temporal structure than white noise that allows them to be perceived easily under scanning conditions [ 18 ].” | content_tags.definitions[1] |
| definition | feature selection — “The purpose of feature selection is to reduce the set of features (in this case, voxels) in a dataset to those most likely to carry relevant information.” | content_tags.definitions[2] |
| definition | categorical perception — “Categorical perception involves mapping a continuum of variation for low-level acoustic features of the stimuli into a discrete number of abstract categories.” | content_tags.definitions[3] |
| definition | tone clouds — “The stimuli were ‘tone clouds’. These are noise-like stimuli with a coarser spectro-temporal structure than white noise that allows them to be perceived easily under scanning conditions” | content_tags.definitions[4] |
| definition | reference tone cloud — “Without the subjects’ knowledge, however, one exemplar of a repeated tone cloud (reference tone cloud, RefTC) remained the same from trial to trial. That is, for the RefTC stimulus not only the same burst of tone cloud was repeated (three times) within a trial, but also the exact same exemplar reoccurred over several trials. The trials consisting of RefTC stimuli were presented randomly among trials of RTC and NTC. A RefTC stimulus was drawn from the same statistical process as any other RTC; the only difference was that the same exemplar of RefTC was presented across different trials.” | content_tags.definitions[5] |
| definition | feature selection — “The purpose of feature selection is to reduce the set of features (in this case, voxels) in a dataset to those most likely to carry relevant information. This is effectively the same as removing voxels most likely to carry noise and is a way of increasing the signal-to-noise ratio.” | content_tags.definitions[6] |
| definition | Categorical perception — “Categorical perception involves mapping a continuum of variation for low-level acoustic features of the stimuli into a discrete number of abstract categories.” | content_tags.definitions[7] |
| definition | template — “However, before a label is assigned, the auditory system must first learn a ‘template’ corresponding to the specific acoustic structure.” | content_tags.definitions[8] |
| key-terms | acoustic patterns, template, tone clouds, reference tone cloud, multi-voxel pattern analysis, support vector machine, feature selection, planum temporale, hippocampus, superior temporal sulcus, entorhinal/perirhinal cortex, parahippocampal cortex, computational hub, categorical perception, unsupervised, memory trace, dprimes, searchlight, cross-validation, fMRI, spectro-temporal |
| t-test t | as written: one tailed t -tests — “Given that we were interested in whether results were significantly above chance, one tailed t -tests were used.” | content_tags.statistics[0] |
| ANOVA F | as written: Repeated measures ANOVAs — “Repeated measures ANOVAs were used to compare accuracy values between regions, and subsequently interrogated using two-tailed paired t -tests.” | content_tags.statistics[1] |
| cross-validation | as written: standard 10-fold cross-validation testing procedure — “The overall classification procedure involved splitting the fMRI data into two segments: a ‘training’ set used to train a classifier with fixed regularization hyperparameter C = 1, and a ‘test’ set used to independently test the classification performance using a standard 10-fold cross-validation testing procedure.” | content_tags.statistics[2] |
| statistical-significance | as written: A threshold of p < 0.05 — “A threshold of p < 0.05 was employed throughout.” | content_tags.statistics[3] |
| support vector machine classification | as written: linear support vector machine (SVM) classifier — “A linear support vector machine (SVM) classifier was created for each ROI.” | content_tags.statistics[4] |
| standard deviation SD | as written: standard deviations (SD) — “Mean volumes (in cubic millimetre, summed across both hemispheres) and standard deviations (SD) for the region of interests (ROIs) were as follows: HC 4188.8 (472.97), EPC 5026.04 (488.71), PHC 1799.29 (256.69), HG 2317.93 (521.93), PT 3967.07 (719.66) and STS 14396.07 (2476.29).” | content_tags.statistics[5] |
| d-prime (signal detection sensitivity) | as written: dprimes — “Post hoc analysis showed that dprimes for the three RefTC were greater than RTC (RefTC-1 > RTC: t 6 = 2.93, p = 0.02; RefTC-2 > RTC: t 6 = 5.12, p = 0.002; RefTC-3 > RTC: t 6 = 2.97, p = 0.02).” | content_tags.statistics[6] |
| multi-voxel pattern analysis | — “Multi-voxel pattern analysis was used to test if the learnt acoustic patterns could be ‘decoded’ from the patterns of activity in the auditory cortex and medial temporal lobe.” | content_tags.statistics[7] |
| linear support vector machine | — “The purpose of feature selection is to reduce the set of features (in this case, voxels) in a dataset to those most likely to carry relevant information. This is effectively the same as removing voxels most likely to carry noise and is a way of increasing the signal-to-noise ratio.” | content_tags.statistics[8] |
| one tailed t -tests | — “The classifier accuracy values for each brain region were compared to chance, which in this case was 33% as we were classifying between three exemplars. Given that we were interested in whether results were significantly above chance, one tailed t -tests were used.” | content_tags.statistics[9] |
| Repeated measures ANOVAs | — “The classifier accuracy values for each brain region were compared to chance, which in this case was 33% as we were classifying between three exemplars. Given that we were interested in whether results were significantly above chance, one tailed t -tests were used. Repeated measures ANOVAs were used to compare accuracy values between regions, and subsequently interrogated using two-tailed” | content_tags.statistics[10] |
| dprimes | — “Analysis of behavioural performance during testing showed a significant effect of stimulus on dprimes ( F 3,18 = 8.98, p = 0.001).” | content_tags.statistics[11] |
| claim type | existence — a second reading reaching the same value; recorded: existence — “As a first step, this study establishes the existence of stabilized representations of sound structure in PT and HC concurrent with auditory learning.” | alternatives.claim_type |
| method | incidental learning of three reference tone-cloud exemplars followed by high-resolution fMRI, with a support vector classifier trained to discriminate them from voxel patterns in six manually segmented regions — a second reading DIFFERING from the recorded value; recorded: fmri — “Multi-voxel pattern analysis was used to test if the learnt acoustic patterns could be ‘decoded’ from the patterns of activity in the auditory cortex and medial temporal lobe.” | alternatives.method |
| phenomenon | unsupervised learning of specific acoustic templates from repeated exposure — a second reading DIFFERING from the recorded value; recorded: implicit learning of random acoustic patterns — “The learning process was fast, robust and unsupervised.” | alternatives.phenomenon |
Full text — tagged
The reference list is not part of the text we hold for this paper. Our copy was extracted from the publisher's page and the bibliography did not survive that extraction, so the article below ends before its references. Nothing has been removed by us.
Abstract Previous behavioural studies have shown that repeated presentation of a randomly chosen acoustic patternterm leads to the unsupervisedterm learning of some of its specific acoustic features. The objective of our study was to determine the neural substrate for the representation of freshly learnt acoustic patternsterm. Subjects first performed a behavioural task that resulted in the incidental learning of three different noise-like acoustic patternsterm. During subsequent high-resolution functional magnetic resonance imaging scanning, subjects were then exposed again to these three learnt patterns and to others that had not been learned. Multi-voxel pattern analysisterm was used to test if the learnt acoustic patternsterm could be ‘decoded’ from the patterns of activity in the auditory cortex and medial temporal lobe. We found that activity in planum temporaleterm and the hippocampusterm reliably distinguished between the learnt acoustic patternsterm. Our results demonstrate that these structures are involved in the neural representation of specific acoustic patternsterm after they have been learnt.
Keywords: acoustic patternsterm, fMRIterm·mth, auditory cortex, multi-voxel pattern analysisterm, hippocampusterm 1. Introduction As humans we are constantly bombarded with sounds, many of which can be identified and assigned a semantic label. However, before a label is assigned, the auditory system must first learn a ‘templateterm’ corresponding to the specific acoustic structure. Although a number of brain imaging studies [ 1 – 4 ] have highlighted the brain system which represents the meaning of sounds, key questions remain as to how the brain learns novel acoustic patternsterm, and whether a specific mechanism exists for the storage of acoustic patternsterm per se .
Behavioural studies have shown that repeated presentation of complex acoustic patternsterm results in the learning of templatesterm. In a series of studies [ 5 – 7 ], subjects were presented with either a 1 s sample of white noise (noise condition, N) or two identical and seamlessly abutting 0.5 s samples of white noise (repeated noise condition, RN). For both RN and N stimuli, samples of white noise were generated anew from trial to trial. However, without this being mentioned to subjects, there was in fact a third type of trials: one particular exemplar of repeated noise was presented over several trials, randomly interspersed throughout an experimental block. These trials (reference repeated noise, RefRN) were thus initially drawn from the same process as RN but, unlike RN, they were acoustically identical across several trials. The task assigned to subjects was to report the presence or the absence of repetition within the noise: after each trial, they pressed one button (‘yes’) if they heard a repetition and another button (‘no’) if not. The main result was that performance was considerably better for the RefRN stimuli than for the RN stimuli. Since the only difference between RefRN and RN was that RefRN was heard over several trials, the improved performance could be attributed to the learning of a templateterm for RefRN. The learning process was fast, robust and unsupervisedterm. Furthermore, learning appeared to be largely implicit: even though this was not systematically quantified, the majority of subjects seemed unaware that the same sound had been presented over different trials. Using a different paradigm, McDermott et al . [ 8 ] showed that when a fixed acoustic patternterm mixed with other randomly chosen acoustic patternsterm is repeatedly presented, segregation of the fixed acoustic patternterm from other patterns could be achieved. This is also consistent with a learning of random patterns through repeated exposure.
Given the suggested importance of repeated exposure for auditory learning, we sought specific brain representations of acoustic patternsterm following repeated exposure. We used high-resolution functional magnetic resonance imaging (fMRIterm·mth) and multi-voxel pattern analysisterm (MVPA). Subjects were first exposed to three acoustic patternsterm, among many other highly similar patterns, to induce learning of the specific patterns. During subsequent fMRIterm·mth scanning, subjects were again exposed to the three learned spectro-temporalterm·key patterns, along with other novel patterns that shared their average spectro-temporalterm·key characteristics. MVPA was used to test if the three exemplars could be ‘decoded’ from the patterns of activity in the brain.
Given evidence for the involvement of primary auditory cortex in storing long-term representations of specific auditory experiences [ 9 – 11 ], we hypothesized that Heschl's gyrusterm (HG), which contains primary auditory cortex, would be involved. We also predicted that areas of non-primary/associative areas of auditory cortex would also be recruited. Specifically, we predicted engagement of planum temporaleterm (PT) which has been hypothesized to be involved in the generation of acoustic ‘templatesterm’ at a stage before semantic processing [ 12 ] and superior temporal sulcusterm (STS), which has been shown to store long-term memories for sounds [ 13 , 14 ]. We also hypothesized that structures in the medial temporal lobes (MTLs) would be crucial. In addition to hippocampusterm (HC), which has been shown to have a role in long-term memory for sounds [ 15 , 16 ], we also speculated that entorhinal/perirhinal cortexterm (EPC) and parahippocampal cortexterm (PHC), both of which receive dense input from the non-primary auditory areas [ 17 ], may be involved.
2. Material and methods (a). Participants Seven healthy subjects (two females, mean age = 22.85 years, s.d. = 1.67 years, range = 20–24 years) with no prior history of neurological and psychiatric disorders participated in the study. All subjects completed a consent form and were paid for their participation.
(b). Stimuli The stimuli were ‘tone cloudsterm’. These are noise-like stimuli with a coarser spectro-temporalterm·key structure than white noise that allows them to be perceived easily under scanning conditions [ 18 ]. The stimuli consisted of multiple brief tones (50 ms) at random frequencies spanning a range from 100 to 10 000 Hz, with random onset times. Specifically, the time-frequency plane was divided into non-overlapping frequency channels and time windows ( figure 1 ). For each cell in the resulting grid, a pure tone was generated with random onset time and random frequency within the cell. This allows the matching of long-term spectrum and temporal envelope for all tone cloudsterm on average. There were two frequency channels per octave and 50 ms per time window. The stimuli were 1.5 s in duration and comprised either three identical and contiguous 0.5 s tone cloudsterm (repeated tone cloudterm, RTC) or three different and contiguous 0.5 s tone cloudsterm (non-repeated tone cloudterm, NTC). Both RTC and NTC stimuli were generated anew for each trial. Without the subjects’ knowledge, however, one exemplar of a repeated tone cloudterm (reference tone cloudterm, RefTC) remained the same from trial to trial. That is, for the RefTC stimulus not only the same burst of tone cloudterm was repeated (three times) within a trial, but also the exact same exemplar reoccurred over several trials. The trials consisting of RefTC stimuli were presented randomly among trials of RTC and NTC. A RefTC stimulus was drawn from the same statistical process as any other RTC; the only difference was that the same exemplar of RefTC was presented across different trials.
Figure 1.
( a ) Schematic of a tone cloudterm stimulus. Each stimulus consisted of brief tone pips at random frequencies (with two channels per octave) and random onset times. ( b ) Illustration of the stimuli and task. The NTC stimuli were formed by concatenating three (0.5 s each) segments of tone cloudterm. The RTC consisted of three repetitions of a single tone cloudterm segment of 0.5 s. Both NTC and RTC were generated anew for each trial. The RefTC also consisted of three repetitions of a single segment of 0.5 s but, importantly, the same stimulus was used for all trials. In the experiment, subjects were presented with a single stimulus and the task was to detect repetitions in the stimulus by pressing one button if repetitions were detected and another if no repetitions were detected.
(c). Training The training paradigm was similar to that employed in [ 5 ]. A single trial consisted of the presentation of a single stimulus (either of RefTC, RTC or NTC category, chosen randomly). After listening to the stimulus, the task of the subject was to detect repetitions in the stimulus by pressing one button if the tone cloudterm stimulus repeated within a trial and pressing another if no repetition was detected ( figure 1 ). A single training block consisted of 20 trials each of RefTC, RTC and NTC stimuli. Three separate blocks of training were used, each consisting of a different exemplar of RefTCs. The RefTC trials were pseudorandomly mixed with trials of RTC and NTC such that RefTC stimuli never occurred on successive trials. All training occurred inside the MRI scanner while it was running in order to create the same conditions during learning as during subsequent testing. Given the repeated exposure to the identical spectro-temporalterm·key structure of the RefTC stimuli (compared to the variable spectro-temporalterm·key structure from trial to trial of the RTC), subjects were expected to form memories of the RefTCs which would be reflected in a better performance on repetition detection for RefTC compared to RTC.
(d). Testing during scanning After training in the MRI scanner, subjects were tested in a single session consisting of 20 trials each for the three trained RefTC exemplars randomly presented with 60 trials of RTC and 120 trials of NTC (which were generated anew). The task was the same as during training: on each trial after listening to a 1.5 s long stimulus, subjects indicated by button presses if the stimulus was repeated. The inter stimulus interval was 3 s. While the subjects were being tested, high-resolution fMRIterm·mth data were continuously acquired. These data were our main focus and were analysed using MVPA. After the acquisition of functional data, a high-resolution structural scan was acquired in the same session. After the MRI scanning, listeners were debriefed by means of a questionnaire. In particular, a question was asked as to whether they thought that any of the sounds recurred during the experiment. One listener said yes, one said some and the remaining five said no.
(e). Magnetic resonance imaging data acquisition All imaging data were acquired on a Siemens 3 T Allegra head only scanner operated with a standard transmit-receive head coil. Functional data (T2 * weighted) were continuously acquired using single-shot high-resolution echo-planar imaging sequence (in-plane resolution = 1.5 × 1.5 mm 2 , field of view = 192 × 192 mm, matrix = 128 × 128, echo time (TE) = 30 ms, asymmetric echo shifted forward by 26 phase-encoding lines, echo spacing = 560 µs). Forty-two interleaved slices, repetition time (TR) 4.2 s, covering auditory cortex (HG, PT), STS and structures in the MTL (HC, PHC and EPC) were acquired. For correction of distortions in the magnetic field, field maps were acquired with a standard manufacturer's double-echo gradient echo field map sequence (TE = 10.0 and 12.46 ms, TR = 1020 ms, matrix size = 64 × 64) with 64 slices covering the whole head (voxel size = 3 mm isotropic). A high-resolution T1-weighted structural MRI scan (voxel size = 1 mm isotropic) was also acquired for each participant after the functional data collection.
(f). Pre-processing of functional magnetic resonance imaging data Pre-processing of the fMRIterm·mth data was carried out using SPM8 ( http://www.fil.ion.ucl.ac.uk/spm/software/spm8/ ). After discarding the first six volumes to allow for magnetic saturation effects, the remaining images were realigned to correct for movement of subjects during scanning. The images were then minimally smoothed with a 3 mm full width at half maximum Gaussian kernel. Each trial was modelled as a separate regressor where the listening time of each trial was modelled as an event and convolved with the canonical haemodynamic response function. Participant-specific movement parametersterm·var were included as regressors of no interest.
(g). Region of interestterm·var segmentation The structural scan of each participant was manually segmented using ITK-SNAP 2.2 [ 19 ] to delineate six regions of interest: HG, PT, STS, HC, PHC and EPC. Examples of segmentations for PT and HC are shown in figure 2 . Segmentation of a structure in each hemisphere was done based on the landmarks and boundaries of that structure in an individual hemisphere.
Figure 2.
Examples of segmented ROIs. Shown on sagittal (upper and middle panel) and coronal (lower panel) sections from a subject chosen at random. HCL, left hippocampusterm; HCR, right hippocampusterm; PTL, left planum temporaleterm; PTR, right planum temporaleterm.
Volumes of HG and PT were defined using the definitions of borders developed in [ 20 , 21 ]. The anterior border of the HG was defined by the complete transverse sulcus (TS), whereas the posterior border was defined using the complete Heschl's sulcus (HS). If there was a repetition of HG, only the anterior gyrus was considered and the posterior was included as part of PT. The postero-medial boundary of the HG was drawn on the axial section by drawing a line from the medial end of TS to the medial end of HS. The lateral boundary of the HG was defined by the termination of HG at the lateral surface of the superior temporal gyrus. The inferior boundary was demarcated in coronal section by noting the stem of HG in that section.
For segmenting PT, the posterior border of HG (Heschl's sulcus) was taken as the anterior border of PT. The markers for the posterior boundary of PT are not well established, because in most cases the posterior end of the sylvian fissure bifurcates into ascending and descending rami, the pattern of which varies across subjects [ 21 ]. There is no consensus on whether the posterior end of PT is limited to the posterior end of the horizontal portion of sylvian fissure [ 22 , 23 ] or extends all the way up to the end-point of the ascending ramus (which could cover a part of the parietal lobe [ 21 ]). In this study, we chose the latter option.
For STS, both upper and lower banks were marked on the coronal section. In subjects where the STS was interrupted by short gyri (‘plis de passage’, [ 24 ]) and ascended to the parietal lobe, all parts (anterior, middle, posterior, ascending anterior and ascending posterior, [ 24 ]) were included in the STS volume.
Hippocampal anatomy was identified using the Duvernoy HC atlas [ 25 ]. The EPC and PHC were segmented according to the protocol described in [ 26 ]. Mean volumes (in cubic millimetre, summed across both hemispheres) and standard deviations (SD)term·stat for the region of intereststerm·var (ROIs) were as follows: HC 4188.8 (472.97), EPC 5026.04 (488.71), PHC 1799.29 (256.69), HG 2317.93 (521.93), PT 3967.07 (719.66) and STS 14396.07 (2476.29).
We initially used standard univariate analyses to interrogate the data, but did not find any significant difference in the mean activity evoked by the three RefTC stimuli in any part of the brain. We therefore focused on using MVPA which we believed would have increased sensitivity in our experimental context.
(h). Multi-voxel pattern analysisterm A linear support vector machine (SVM) classifierterm·stat was created for each ROI. Each classifier was trained on a portion of the fMRIterm·mth data relating to the three different exemplars of RefTCs and then tested on an independent set of instances of these exemplars.
We used a standard MVPA procedure that has been described in detail elsewhere [ 27 , 28 ] (for an in depth review, see [ 29 ]). The overall classification procedure involved splitting the fMRIterm·mth data into two segments: a ‘training’ set used to train a classifier with fixed regularization hyperparameter C = 1, and a ‘test’ set used to independently test the classification performance using a standard 10-fold cross-validationterm·key testing procedure. This therefore generated 10 sets of SVM training and test sets that produced overall classification accuracyterm·var from the proportion of correct classification ‘guesses’ across all 10-folds of the cross-validationterm·key. The classification was performed using the LIBSVM implementation [ 30 ]. Prior to multivariate classification, feature selectionterm [ 31 ] was performed on the data from the training set. This was conducted using a standard multivariate searchlightterm·key strategy within the given ROI.
(i). Feature selectionterm for multi-voxel pattern analysisterm The purpose of feature selectionterm is to reduce the set of features (in this case, voxels) in a dataset to those most likely to carry relevant information. This is effectively the same as removing voxels most likely to carry noise and is a way of increasing the signal-to-noise ratio. This was conducted using a standard multivariate searchlightterm·key strategy within the given ROI. For a given voxel, we first defined a small sphere with a radius of three voxels centred on a given voxel [ 27 , 32 , 33 ]. Note that the spheres were restricted so that only voxels falling within the given ROI were included. Therefore, the shape of the sphere and the number of voxels within it varied depending on the proximity to the ROI's borders. This procedure then allowed the selection of the searchlightterm·key voxel set that contained the greatest degree of decoding information within the training dataset. Using this voxel subset, the SVM classifier was trained to discriminate between the three RefTCs using the ‘training’ dataset, and tested on the independent ‘test’ dataset.
Standard SVMs are binary classifiers that operate on two-class discrimination problems, whereas our data involved a three-class problem (i.e. three exemplars). The SVM can, however, be arbitrarily extended to work in cases in which there are more than two classes. Typically, this is done by reducing the single multiclass problem into multiple binary classification problems that can be solved separately and then recombined to provide the final class prediction [ 34 ]. We used the well-established error correcting output codes approach [ 35 ] and computing of the Hamming distance [ 27 , 32 , 36 ]. The classifier accuracy values for each brain region were compared to chance, which in this case was 33% as we were classifying between three exemplars. Given that we were interested in whether results were significantly above chance, one tailed t -teststerm·stat were used. Repeated measures ANOVAsterm·stat were used to compare accuracy values between regions, and subsequently interrogated using two-tailed paired t -tests. A threshold of p < 0.05 was employed throughout. Since the training phase and the testing phase used different paradigms (during training separate RefTCs were presented in different blocks and in testing the three learned exemplars of RefTCs were presented in the same block), the fMRIterm·mth data from the training phase were not analysed, and we focused on our main question of where the learned RefTCs were represented.
3. Results (a). Behavioural performance Behavioural performance during training and testing is shown in figure 3 . There was a significant effect of stimulus on performance during training ( F 3,18 = 7.08, p = 0.002). Post hoc analysis showed that dprimesterm for the three RefTC were greater than RTC (RefTC-1 > RTC: t 6 = 2.93, p = 0.02; RefTC-2 > RTC: t 6 = 5.12, p = 0.002; RefTC-3 > RTC: t 6 = 2.97, p = 0.02). Performance on the three RefTC did not differ significantly. Analysis of behavioural performance during testing showed a significant effect of stimulus on dprimesterm ( F 3,18 = 8.98, p = 0.001). Post hoc comparison revealed better performance on all the RefTC stimuli compared to RTC (RefTC-1 > RTC, t 6 = 3.96, p = 0.007; RefTC-2 > RTC, t 6 = 4.30, p = 0.005; RefTC-3 > RTC, t 6 = 5.43, p = 0.001). Thus, consistent with previous findings [ 5 , 6 , 37 ], the behavioural data show learning of spectro-temporalterm·key patterns that are repeatedly presented during the course of the experiment. The novel question we then addressed was where in the brain these RefTCs were represented.
Figure 3.
Behavioural performance during ( a ) training and ( b ) testing. dprimesterm (mean ± s.e. of mean) are plotted for the three exemplars that were repeated during the course of experiment (RefTC) and the exemplars that were the same structure as RefTC but were generated anew in each trial (RTC). RefTC-1, repeated reference exemplar 1; RefTC-2, repeated reference exemplar 2; RefTC-3, repeated reference exemplar 3; RTC, repeated stimulus. (* p < 0.05.) (b). Multi-voxel pattern analysisterm Using MVPA we examined whether it was possible to discriminate which of the three RefTC stimuli was being heard solely from the pattern of activity across voxels in our ROIs. For each ROI, a linear SVM classifier was first trained on a portion of the fMRIterm·mth data relating to the three RefTC stimuli and then tested on an independent set of trials of these stimuli. If information was present in the patterns of fMRIterm·mth activity that would allow for successful discrimination between the three stimuli, then the classifier would produce a result that was significantly above chance (33%). There was no statistically significant difference between the classification accuracies of left and right hemispheres (HC: t 6 = 0.465, p = 0.658; EPC: t 6 = −1.666, p = 0.147; PHC: t 6 = −0.511, p = 0.628; HG: t 6 = −1.384, p = 0.216; PT: t 6 = −1.532, p = 0.177; STS: t 6 = 0.789, p = 0.460) and therefore the reported data are collapsed across hemispheres.
Classification accuracyterm·var for the six ROIs is shown in figure 4 . Only two regions, HC and PT, showed performance above chance: (HC: t 6 = 2.711, p = 0.018; EPC: t 6 = 0.188, p = 0.43; PHC: t 6 = 0.117, p = 0.46; HG: t 6 = 0.620, p = 0.28; PT: t 6 = 5.106, p = 0.001; STS: t 6 = 1.715, p = 0.07). A repeated measures ANOVAterm·stat showed a significant effect for region ( F 5,30 = 3.271, p = 0.018), which was driven by more information being present in the HC when compared with HG (HC > HG: t 6 = 2.765, p = 0.033) as well as more information present in PT when compared with HG, EPC and PHC (PT > HG: t 6 = 4.182, p = 0.006; PT > EPC: t 6 = 3.059, p = 0.022; PT > PHC: t 6 = 3.291, p = 0.017). Classification accuracyterm·var in the STS was almost significant (see above) and did not differ significantly from either PT ( t 6 = 1.91, p = 0.11) or HC ( t 6 = 0.25, p = 0.81).
Figure 4.
Results of the MVPA analysis. The mean (±s.e. of mean) classifier performance (collapsed across both hemispheres) is shown. Chance was 33% (marked by the dotted line). HC, hippocampusterm; EPC, entorhinal/perirhinal cortexterm; PHC, parahippocampal cortexterm; HG, Heschl gyrus; PT, planum temporaleterm; STS, superior temporal sulcusterm. (* p < 0.05.) To ensure that the classifiers were unbiased and that classification was based on unique representations for each of the three RefTCs, we ran a control analysis in which we randomly shuffled (from trial to trial) the labels of the three RefTCs. The classifiers were trained and tested as above. As expected, the performance of the classifier dropped to chance level for all the brain regions (HC: t 6 = −0.341, p = 0.628; EPC: t 6 = 0.735, p = 0.245; PHG: t 6 = 1.047, p = 0.168; HG: t 6 = 0.392, p = 0.354; PT: t 6 = 0.109, p = 0.459; STS: t 6 = −0.147, p = 0.557).
To further confirm that classification of RefTCs was based on stable representations (owing to the repeated presentation of the same exemplars), we ran a second control analysis. In this analysis, we divided the 60 RTC trials randomly into three classes and performed the same classification analysis. Since a different exemplar is presented in every trial of RTC, representation of RTC changes from trial to trial. It is therefore expected that the classifier should not be able to classify the RTC stimuli significantly better than chance. As predicted, the classifier performance was at chance for all ROIs, (HC: t 6 = −0.058, p = 0.955; EPC: t 6 = −0.421, p = 0.689; PHG: t 6 = 0.953, p = 0.377; HG: t 6 = 1.365, p = 0.221; PT: t 6 = −0.174, p = 0.868; STS: t 6 = −0.316, p = 0.763).
4. Discussion As in AguspersName et al. [ 5 ], the behavioural data in our study revealed better performance on repeated exemplars compared to performance on non-repeated exemplars confirming learning of acoustic patternsterm. Importantly, the repeatedly presented stimuli (RefTCs) and the non-repeated stimuli (RTC) were well balanced with respect to acoustic parametersterm·var. The key difference between the two conditions was in terms of exposure over the course of the experiment. Subjective reports from the participants after the experiment showed that most were not aware of any stimulus re-occurring during the experiment, suggesting that in some cases the learning occurred implicitly.
We found that patterns of activity across voxels in PT, but not in HG, could distinguish between the three learned acoustic patternsterm. We speculate that this might reflect a type of representation within PT that is not a simple representation of acoustic patternterm but a more refined representation that requires interaction with HC (see below). The availability of pattern specific information in the PT is consistent with a role as a computational hubterm [ 12 ], where spectro-temporalterm·key patterns are matched with learned patterns that are stored beyond the auditory cortex. Our results showing the involvement of auditory cortex in the storage of long-term representations of stimuli without semantic association is consistent with a recent magnetoencephalography study [ 37 ] that showed different phase patterns for different noise exemplars that had been learnt, suggesting a specificity of the representation for a given pattern. However, compared to the results in [ 37 ], our results are more specific with respect to the areas of auditory cortex involved in storing pre-semantic-specific templatesterm.
Our results show that HC is involved in representations of RefTC stimuli that are unique to each exemplar of RefTC. Although HC is known to be involved in processing of complex and meaningful sounds [ 15 , 38 ], to the best of our knowledge, our study is the first to demonstrate the encoding of noise-like acoustic patternsterm in the HC. There are some studies in the visual domain [ 39 ] which showed sensitivity of the HC to changes in low-level perceptual features (e.g. change in font size of displayed letters). The results of our study further extend these results by showing that the HC is not only sensitive to low-level features of stimuli that are explicitly recalled, but also constructs representations that are specific to acoustic features which are learned implicitly.
Our results also showed that the classifier performance was close to significant ( p = 0.07) for the STS region. This lack of significance may be owing to the low power of our study, so a role for STS in the learning of novel acoustic patternsterm cannot be ruled out. The STS is a broad region which has been implicated in a wide range of unimodal and multimodal functions (for review, see [ 40 ]). From the auditory perception point of view, converging evidence from neuroimaging studies show that STS is involved in categorical perceptionterm of speech [ 13 ] and non-speech [ 14 , 41 ] auditory stimuli. Categorical perceptionterm involves mapping a continuum of variation for low-level acoustic features of the stimuli into a discrete number of abstract categories. Furthermore, the connectivity analysis [ 42 ] between the HG, PT and STS shows that representations of acoustic features in HG and PT are relayed to the STS for object like representations. In the context of current study, it is therefore likely that although each of the three RefTC's has a unique representation in the PT, these representations may have been further abstracted and instantiated in the STS.
We considered whether our results might be explained by sensory or perceptual representations in PT and/or HC. The existence of sensory representations of fine spectro-temporalterm·key features that occurs in non-primary auditory cortex and HC but not in primary auditory cortex is unlikely. Mapping of the different perceptual timbre of different exemplars could occur in non-primary auditory cortex, but there is no precedent for any effect of manipulating spectro-temporalterm·key structure and timbre on hippocampal activity [ 43 – 46 ]. The most parsimonious explanation for our data is a unique memory traceterm in PT and HC that is formed for each of the three RefTCs, although more work will be needed to probe this point further. It will also be important to examine another issue in the future. Our learning phase was optimized to facilitate acquisition of the three RefTCs; consequently, the fMRIterm·mth data acquired during this phase were not intended (nor were they suitable) for analysis using MVPA, because our focus was solely on where the learnt RefTCs were represented. Further studies could investigate the learning over time of RefTCs to see if PT and HC (and/or other brain regions) are implicated.
Structural connectivity between human auditory cortex and the MTLs is not completely understood. However, the connectivity pattern between auditory cortex and the MTL of monkeys [ 47 ] (for review, see [ 17 ]) shows that belt and parabelt, but not the core, of the auditory cortex have direct projections to the entorhinal cortex. The human PT we examined in our experiment is a non-core area containing homologues of primarily auditory belt (and possibly auditory parabelt) cortex [ 48 ]. Based on the non-human primate work, therefore, it appears likely that PT and HC are reciprocally connected. The connectivity between PT and HC offers good grounds for proposing that PT and HC might form a connected system that allows specific representations of learnt spectro-temporalterm·key patterns. The purpose of such a system may be to transform the rich pattern of activity for complex sounds expressed in HG into sparser representationsterm, more amenable to long-term memory storage. This raises questions about the dynamics of this system, as behavioural data show that learning occurs rapidly. Further work is required to determine how many repetitions of each exemplar are required during the training phase before a stable representation is built in the PT and HC. A further question concerns the direction of causal influences of the PT and HC on each other that occur during the construction of stable representations. As a first step, this study establishes the existence of stabilized representations of sound structure in PT and HC concurrent with auditory learning.
The study was approved by the local research ethics committee.
Data accessibility The procedure we have agreed with the UCL Ethics Committee is that any request for the data needs to be initiated from an appropriate source (usually a fellow scientist) after which we will make the anonymized data available to him or her.
Funding statement This work was funded by the Wellcome Trust (WT091681MA, WT084218).