Pith. sign in

REVIEW 3 major objections 5 minor 49 references

Local and non-local dependency learning and emergence of rule-like representations in speech data by Deep Convolutional Generative Adversarial Networks

T0 review · 3 major / 5 minor · reviewed 2026-08-27 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A deep convolutional GAN trained on raw speech encodes a morphological prefix in one latent variable, and interpolating a second variable makes the prefix vowel follow the trigger—an approximation of a non-local phonological rule.

desk verdict The z16 prefix result is solid, but the paper's stronger claim that a non-local harmony rule emerges from the latent space is not yet supported because z17 may directly control both vowels. read the letter →

arxiv 2009.12711 v2 pith:TZFY6R4O submitted 2020-09-27 cs.CL cs.LG

Gašper Beguš This is my paper · ORCID
classification cs.CLcs.LG
keywords neuralnetworksbehavioralexperimentsmachinelearningbiasesspeechmorphologyvowelharmonylatentspace
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that a deep convolutional GAN trained from raw audio on a small artificial lexicon learns both local and non-local phonological dependencies and stores at least some of what it learns as discrete, single-variable codes in its latent space. The central result is that the presence of a morphological prefix is controlled by one latent variable, and the front/back quality of the triggering vowel by another; interpolating the vowel variable while forcing the prefix produces an automatic front-to-back change in the prefix vowel, an approximation of a non-local vowel-harmony rule. The paper also reports that local devoicing processes are learned far more reliably than non-local harmony, matching human behavioral data and the typological preference for locality, and that the model remains informative even with only 270 training items. If these results hold, they suggest that rule-like morphophonological behavior can emerge from an unsupervised convolutional network without language-specific machinery or pre-discretized symbolic input.

What carries the argument

The load-bearing object is the generator's 100-dimensional latent vector $\mathbf{z}$, with each component drawn uniformly from $(-1,1)$ during training. The paper identifies meaningful directions in this space by fitting penalized logistic regressions of human transcriptions of generated audio on the latent values, then tests those directions by setting individual variables far outside the training range (e.g. $z_{16} = \pm 4.5$) and interpolating $z_{17}$ from $-6$ to $6$. A second methodological device is keeping the latent vector constant while generating from networks at different training steps, which lets the author watch the network repair harmony and devoicing violations over time. The claim that a single variable 'means' prefix or frontness rests on these paired regression-and-generative tests.

What would settle it

Take the trained generator, hold $z_{16}$ at $-2.5$, generate the 13-step $z_{17}$ interpolation 60 times as in the paper, and have a second annotator who is blind to the latent values transcribe the 780 outputs, or run an automatic formant tracker on the prefix vowel. If prefix-vowel quality no longer follows the triggering vowel, or if the $z_{16}$ prefix effect drops from 100% versus 1%, the central claim is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that approximations of symbolic phonological rules are present in a deep convolutional GAN trained on raw speech. Penalized logistic regression on 500 generated outputs shows a steep drop in coefficients after $z_{16}$ for prefix presence and after $z_{17}$ for vowel frontness, indicating that each property is encoded by a single latent variable. Setting $z_{16}$ to $-4.5$ yields a prefix in 100/100 samples; setting it to $4.5$ yields a prefix in 1/100. When $z_{16}$ is fixed at $-2.5$ and $z_{17}$ is interpolated from $-6$ to $6$, the triggering vowel shifts from front to back and the prefix vowel follows it, staying harmonious in the majority of outputs—the network's analogue of a vowel-harmony rule. The paper is explicit that this is probabilistic rather than categorical, with harmony rates around 77% overall and a dip in the transition region.

Load-bearing premise

The load-bearing premise is that the annotations of prefix presence, vowel frontness, and harmony in generated audio are accurate; they come from a single, unblinded human annotator, so biased judgments could manufacture the apparent single-variable control.

Editorial extensions

If this is right

  • If correct, the method gives a way to read discrete linguistic structure out of a continuous neural network: fit regression, set the variable, and watch the output change, making internal representations observable.
  • The strong local/non-local error gap implies that a domain-general learning bias need not be built in; a deep convolutional learner can produce the typological preference for locality from raw acoustics alone.
  • Because 270 items sufficed, researchers can run matched computational and human artificial-grammar-learning experiments on identical small stimuli, as the paper does.
  • The constant-latent-space technique across training steps can reveal the order in which rules are acquired, and should be used to avoid ceiling effects when comparing process learnability.
  • An approximation to a morphophonological rule can be elicited from an unsupervised model with no language-specific parameters, bearing on the symbolism-versus-connectionism debate.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A blind test is the natural next step and is our inference, not the paper's: ask a second annotator who does not know which $z$ values were used to transcribe the same generated samples, or use an automatic formant classifier on the prefix vowel; if the $z_{16}$/$z_{17}$ separation disappears, the result is an annotation artifact.
  • If single-variable discretization holds for other morphological patterns, the same interpolation diagnostic could reveal discrete latent directions for reduplication, tone, or affix order in GANs trained on other artificial languages.
  • The observed dip in harmony during the front-to-back transition suggests a testable prediction for human learners: accuracy should be lowest when the triggering vowel is acoustically intermediate, not simply lower overall for non-local processes.
  • The paper's comparison between network and human error rates is coarse; a stronger extension would play the network's own generated items to human listeners in a two-alternative forced choice, testing whether the same acoustic cues drive both judgments.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. This paper trains a WaveGAN/DCGAN on 270 raw audio nonce words containing prefixation, vowel harmony, and local devoicing patterns, and uses Lasso regression on 100 latent variables to identify single latent dimensions controlling prefix presence (z16) and V2 frontness (z17). By setting z16 to -2.5 and interpolating z17 from -6 to 6, the author reports that generated prefix vowels remain harmonious with V2, which is interpreted as an emergent rule-like non-local process. The paper also reports that local devoicing is learned with 1.8% error versus 23.2% harmony error, and compares these rates to an artificial grammar learning experiment. The central claims are that GANs trained on small raw-audio data can encode a morphological unit in a single latent variable and approximate symbolic rule computation through latent interactions.

Significance. If correct, the result is significant: it would demonstrate that a DCGAN with no language-specific components learns a discrete morphological representation and a non-local harmony interaction from raw audio, and that local processes are easier than non-local ones in a way that parallels human learning. The strongest element is the generative test for z16: at -4.5, 100/100 outputs contain a prefix, versus 1/100 at +4.5, and because these values lie outside the training range and were not used to fit the regressions, this is a genuine extrapolative test rather than a restatement of the fitted model. The data and trained models are shared via OSF, which supports reproducibility. The z17 harmony result is more fragile because it rests on single-annotator labels and does not exclude a global frontness confound.

major comments (3)
  1. [Section 3.5 (Figure 6, Table A.13)] The claim that z17 is causally specific to V2 is not established. z17 was selected by regressing V2 frontness on the latent space using the 500 outputs from Section 3.3, but in those outputs prefix-vowel frontness is strongly correlated with V2 frontness (77.1% harmonious, Table 2). A z17 that directly sets a global vowel-frontness property, or specifically the prefix vowel, would therefore also be the best predictor of V2 frontness. Conditioning on annotated V2 frontness in the harmony model (Table A.13) does not remove a direct z17-to-prefix-vowel path, and no test is reported that manipulates z17 while keeping the prefix absent or that includes prefix-vowel frontness as a covariate. Without such a control, the observed covariation is equally compatible with a non-local interaction and with a single latent dimension that globally controls vowel quality.
  2. [Section 3.3 and Section 3.5] All quantitative evaluations of generated audio rest on the author's own transcription and acoustic annotation, as stated in Section 3.3: 'All acoustic analyses are performed in Praat by the author.' The harmony, frontness, and devoicing labels used in both Lasso regressions and the generative tests are single-annotator judgments, with no inter-annotator reliability, no blinding to the manipulated latent variables, and no independent acoustic classifier. If the annotator's expectations influence the labels (for example, in judging whether a forced prefix vowel is [E] or [O]), the z17 result could be partly an artifact of labeling. The paper should provide a second annotation on a subset, report agreement (e.g., Cohen's kappa), and ideally blind the annotator to z16/z17 values.
  3. [Section 3.3 (Figure 4) and Section 3.5 (Figure 6)] The 'substantial drop' in Lasso regression estimates is used as evidence for single-variable discretized encoding, but no statistical criterion or null model is given. With 100 correlated predictors and a single checkpoint, a steep drop can occur by chance or due to the regularization path; the paper should report the cross-validated lambda, coefficient values, stability across random seeds, or a permutation baseline. This does not undermine the z16 generative test (100% vs 1%), but it weakens the more general claim that the network 'discretizes' representations into single latent variables.
minor comments (5)
  1. [Section 3.5] There is a typo: 'To achieve this affect' should be 'To achieve this effect'.
  2. [Section 3.3] The choice of the 20990-step checkpoint for all quantitative analyses is justified post hoc from the exploratory analysis in Section 3.2; since multiple checkpoints were inspected, the paper should acknowledge the multiple-comparisons issue or report the same tests at at least one other checkpoint.
  3. [Section 3.5] For the 60 sets of 13 generated samples, it would be useful to state explicitly whether each set uses a different random draw of the remaining 98 latent variables and whether z16 was fixed to -2.5 for all samples; this affects how the by-trajectory random effects should be interpreted.
  4. [Table A.10] The contrast labels 'mean vs. back' and 'mean vs. V-' are unclear; use explicit sum-coded contrast names such as 'frontness (front vs back)' and 'prefix (V- vs VN-)'.
  5. [Figure 10] The human and network panels use different model types (mixed-effects logistic vs plain logistic) and different tasks; the text already acknowledges the comparison is superficial, but the figure would benefit from a note to prevent readers from over-interpreting the visual alignment of percentages.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the latent-variable identification and interpolation tests use new outputs and out-of-training-range values, so the central claims do not reduce to their inputs.

full rationale

The paper's derivation chain is: annotate 500 generated outputs, fit Lasso regressions to select z16 (prefix presence) and z17 (V2 frontness), then generate fresh outputs with z16 and z17 set to extreme values and test whether prefixation and harmony track the manipulation. The generative tests are not re-statements of the fitted regressions: z16 is set to ±4.5 and z17 is interpolated from -6 to 6, values that never occur in the training distribution (z ~ U(-1,1)) and were not used in the Lasso fits. The observed 100%/1% prefix rates for z16 and the significant front-to-back shift in V2 for z17 are extrapolation results, not fitted quantities renamed as predictions. Likewise, the harmony test in Section 3.5 annotates newly generated interpolated outputs and fits a separate GAMM for harmony; harmony was not the dependent variable in the regression that selected z17, so the 'automatic' prefix-vowel change is not the fitted parameter. The self-citations to Beguš (2020b) import the general latent-regression-and-interpolation technique, but the present paper runs its own in-paper generative tests, so the citation is not the sole load-bearing support for the central claim. The legitimate concerns—single-annotator labels and the possibility that z17 directly controls both V2 frontness and prefix-vowel frontness—are confounds and robustness issues rather than definitional circularity; they do not exhibit a specific reduction of the prediction to the input by construction. Under the required standard of quoting an equation or fitted parameter that is equivalent to the claimed result, no circular step is established.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims rest on: (1) author annotation of outputs, (2) the imported Beguš (2020b) identification technique, (3) extrapolation of z-values outside the training range, and (4) a visual 'drop' criterion for single-variable encoding. No new entities are postulated, and no parameters are fitted to the target conclusion beyond the checkpoint and manipulation values described above.

free parameters (3)
  • Training checkpoint for main analyses = 20990 steps (plus 7453, 9740, 14900 for progression)
    The checkpoint was chosen post hoc to maximize acoustic clarity and because disharmony appeared repaired at 20990 steps (Sections 3.1 and 3.3); all results in Sections 3.3 to 3.5 depend on this selection.
  • Latent manipulation values = z16 = -2.5; z17 from -6 to 6
    These values lie outside the training range (-1, 1) and were chosen from regression estimates without a principled criterion (Section 3.5).
  • Annotation sample sizes = 500 outputs; 60 trajectories of 13 samples (780 outputs)
    The sample sizes for the regressions and generative tests are arbitrary choices not justified by power analysis (Sections 3.3 and 3.5).
assumptions (4)
  • domain assumption Generated audio can be reliably transcribed into IPA segments, with prefix presence, vowel frontness, voicing, and harmony identifiable from spectrograms and listening.
    Entered throughout Sections 3.3 to 3.5; no inter-annotator agreement or validation against an objective acoustic classifier is provided.
  • domain assumption The latent-variable identification technique from Beguš (2020b) is valid for this dataset and setup.
    The paper imports the Lasso-based variable identification and out-of-range manipulation method from prior work (Sections 2.1 and 3.3) without re-deriving or validating it here.
  • domain assumption Linear extrapolation of latent variables beyond the training range (-1, 1) exposes the same learned feature dimensions in a monotonic way.
    z16 is set to -2.5 and z17 is swept from -6 to 6 (Section 3.5); this assumes the learned direction generalizes outside the support of the training distribution.
  • ad hoc to paper The 'substantial drop' in Lasso regression estimates constitutes evidence of a single-variable discretized encoding.
    Sections 3.3 and 3.5 infer single-variable encoding from the visual drop in estimates (z16, z17); no statistical test or model comparison with and without that variable is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Local and non-local dependency learning and emergence of rule-like representations in speech data by Deep Convolutional Generative Adversarial Networks." pith.science (2026). https://pith.science/paper/TZFY6R4O

@misc{pith2026200912711,
  author       = {Pith},
  title        = {Pith review of: Local and non-local dependency learning and emergence of rule-like representations in speech data by Deep Convolutional Generative Adversarial Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TZFY6R4O}},
  note         = {Machine review of arXiv:2009.12711}
}
read the original abstract

This paper argues that training GANs on local and non-local dependencies in speech data offers insights into how deep neural networks discretize continuous data and how symbolic-like rule-based morphophonological processes emerge in a deep convolutional architecture. Acquisition of speech has recently been modeled as a dependency between latent space and data generated by GANs in Begu\v{s} (2020b; arXiv:2006.03965), who models learning of a simple local allophonic distribution. We extend this approach to test learning of local and non-local phonological processes that include approximations of morphological processes. We further parallel outputs of the model to results of a behavioral experiment where human subjects are trained on the data used for training the GAN network. Four main conclusions emerge: (i) the networks provide useful information for computational models of speech acquisition even if trained on a comparatively small dataset of an artificial grammar learning experiment; (ii) local processes are easier to learn than non-local processes, which matches both behavioral data in human subjects and typology in the world's languages. This paper also proposes (iii) how we can actively observe the network's progress in learning and explore the effect of training steps on learning representations by keeping latent space constant across different training steps. Finally, this paper shows that (iv) the network learns to encode the presence of a prefix with a single latent variable; by interpolating this variable, we can actively observe the operation of a non-local phonological process. The proposed technique for retrieving learning representations has general implications for our understanding of how GANs discretize continuous speech data and suggests that rule-like generalizations in the training data are represented as an interaction between variables in the network's latent space.

Figures

Figures reproduced from arXiv: 2009.12711 by the authors.

Figure 1
Figure 1. The GAN architecture schematized from Goodfellow et al. (2014); Radford et al. (2015); Donahue et al. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. (left) Waveforms and spectrograms (0–4000 Hz) of three generated samples with the same values of latent variables at three different training steps. (right) Waveforms and spectrograms (0–8000 Hz) of four generated samples with the same values of latent variables at four training steps showing devoicing, change of place of articulation, and occlusion. output, which means the output now conforms to the devoicing rule … view at source ↗
Figure 3
Figure 3. Waveforms and spectrograms (0–8000 Hz) of three outputs illustrating changes in outputs with the same [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Absolute Lasso logistic regression estimates of a model with presence of the prefix as the dependent [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Waveforms and spectrograms (0–8000 Hz) of three outputs of the Generator network trained after 20990 [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: (a) Absolute Lasso logistic regression estimates of a model with presence of front triggering vowels V2 as the dependent variable and values of 100 z-variables as independent predictors. The estimates are sorted in reversed order. (b) Absolute Lasso logistic regression…
Figure 7
Figure 7. Figure 7: (a) Fitted values and 95% CIs of a generalized additive mixed effects logistic regression model with the front vs. back triggering vowel (V2) value as the dependent variable and thin-plate smooths for values of z17 as the independent variable (with random smooths for e…
Figure 8
Figure 8. Figure 8: Waveforms and spectrograms (0–5000 Hz) of outputs with interpolated values of [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Experimenal design (from Beguˇs 2020a) in the Experigen interface (Becker and Levine, 2013); artwork [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: (a) Estimates of the linear mixed effects logistic regression model with harmonious responses of human subjects in the behavioral experiment as successes and vowel frontnesss and prefix identity as the independent variables with their interaction. (b) Estimates and 95…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages

  1. [1]

    Investigating under and overfitting in wasserstein generative adversarial networks

    Adlam, B., Weill, C., Kapoor, A., 2019. Investigating under and overfitting in wasserstein generative adversarial networks

    work page 2019

  2. [2]

    Alishahi, A., Barking, M., Chrupa a, G., Aug. 2017. Encoding of phonology in a recurrent neural model of grounded speech. In: Proceedings of the 21st Conference on Computational Natural Language Learning ( C o NLL 2017). Association for Computational Linguistics, Vancouver, Canada, pp. 368--378. ://www.aclweb.org/anthology/K17-1037

    work page 2017

  3. [3]

    W asserstein generative adversarial networks

    Arjovsky, M., Chintala, S., Bottou, L., 06--11 Aug 2017. W asserstein generative adversarial networks. In: Precup, D., Teh, Y. W. (Eds.), Proceedings of the 34th International Conference on Machine Learning. Vol. 70 of Proceedings of Machine Learning Research. PMLR, International Convention Centre, Sydney, Australia, pp. 214--223. ://proceedings.mlr.press...

    work page 2017

  4. [4]

    Fitting linear mixed-effects models using lme4

    Bates, D., M \"a chler, M., Bolker, B., Walker, S., 2015. Fitting linear mixed-effects models using lme4 . Journal of Statistical Software 67 (1), 1--48

    work page 2015

  5. [5]

    Experigen -- an online experiment platform

    Becker, M., Levine, J., 2013. Experigen -- an online experiment platform. ://becker.phonologist.org/experigen

    work page 2013

  6. [6]

    Ciwgan and fiwgan: Encoding information in acoustic data to model lexical learning with generative adversarial networks

    Begu s , G., 2021. Ciwgan and fiwgan: Encoding information in acoustic data to model lexical learning with generative adversarial networks. Neural Networks 139, 305--325. ://www.sciencedirect.com/science/article/pii/S0893608021001052

    work page 2021

  7. [7]

    Distinguishing cognitive from historical influences in phonology, submitted ms., UC Berkeley

    Begu s , G., 2020 a . Distinguishing cognitive from historical influences in phonology, submitted ms., UC Berkeley

    work page 2020

  8. [8]

    Frontiers in Astronomy and Space Sciences , keywords =

    Begu s , G., 2020 b . Generative adversarial phonology: Modeling unsupervised phonetic and phonological learning with neural networks. Frontiers in Artificial Intelligence 3, 44. ://www.frontiersin.org/article/10.3389/frai.2020.00044

    work page arXiv 2020

Show all 49 references
  1. [9]

    The phonological mind

    Berent, I., 2013. The phonological mind. Trends in Cognitive Sciences 17 (7), 319 -- 327. ://www.sciencedirect.com/science/article/pii/S1364661313001034

    2013

  2. [10]

    Praat: doing phonetics by computer [computer program]

    Boersma, P., Weenink, D., 2015. Praat: doing phonetics by computer [computer program]. version 5.4.06. Retrieved 21 February 2015 from http://www.praat.org/

    2015

  3. [11]

    S., Wilson, H

    Bond, Z. S., Wilson, H. F., 1980. /s/ plus stop clusters in children's speech. Phonetica 37 (3), 149--158. ://www.karger.com/DOI/10.1159/000259988

    1980 doi

  4. [12]

    Generative Adversarial Networks with Python: Deep Learning Generative Models for Image Synthesis and Image Translation

    Brownlee, J., 2019. Generative Adversarial Networks with Python: Deep Learning Generative Models for Image Synthesis and Image Translation. Machine Learning Mastery

    2019

  5. [13]

    Infogan: Interpretable representation learning by information maximizing generative adversarial nets

    Chen, X., Duan, Y., Houthooft, R., Schulman, J., Sutskever, I., Abbeel, P., 2016. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. In: Lee, D. D., Sugiyama, M., Luxburg, U. V., Guyon, I., Garnett, R. (Eds.), Advances in Neur...

    2016

  6. [14]

    The Sound Pattern of E nglish

    Chomsky, N., Halle, M., 1968. The Sound Pattern of E nglish. Harper & Row, New York

    1968

  7. [15]

    Vector-Quantized Autoregressive Predictive Coding

    Chung, Y.-A., Tang, H., Glass, J., 2020. Vector-Quantized Autoregressive Predictive Coding . In: Proc. Interspeech 2020. pp. 3760--3764. ://dx.doi.org/10.21437/Interspeech.2020-1228

    2020 doi

  8. [16]

    J., Puckette, M

    Donahue, C., McAuley, J. J., Puckette, M. S., 2019. Adversarial audio synthesis. In: 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, pp. 1--16). ://openreview.net/forum?id=ByMVTsR5KQ

    2019

  9. [17]

    Unsupervised acoustic unit discovery for speech synthesis using discrete latent-variable neural networks

    Eloff, R., Nortje, A., van Niekerk, B., Govender, A., Nortje, L., Pretorius, A., Biljon, E., van der Westhuizen, E., Staden, L., Kamper, H., 09 2019. Unsupervised acoustic unit discovery for speech synthesis using discrete latent-variable neural networks. In: Proc. Interspeech...

    2019

  10. [18]

    The privileged status of locality in consonant harmony

    Finley, S., 2011. The privileged status of locality in consonant harmony. Journal of Memory and Language 65 (1), 74 -- 83. ://www.sciencedirect.com/science/article/pii/S0749596X11000192

    2011

  11. [19]

    Testing the limits of long-distance learning: Learning beyond a three-segment window

    Finley, S., 2012. Testing the limits of long-distance learning: Learning beyond a three-segment window. Cognitive Science 36 (4), 740--756. ://onlinelibrary.wiley.com/doi/abs/10.1111/j.1551-6709.2011.01227.x

    2012

  12. [20]

    S., Lamel, L., M Fisher, W., Fiscus, J., S

    Garofolo, J. S., Lamel, L., M Fisher, W., Fiscus, J., S. Pallett, D., L. Dahlgren, N., Zue, V., 11 1993. Timit acoustic-phonetic continuous speech corpus. Linguistic Data Consortium

    1993

  13. [21]

    D., 1995

    Gaskell, M., Hare, M., Marslen-Wilson, W. D., 1995. A connectionist model of phonological representation in speech perception. Cognitive Science 19 (4), 407 -- 439. ://www.sciencedirect.com/science/article/pii/0364021395900071

    1995

  14. [22]

    Generative adversarial nets

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y., 2014. Generative adversarial nets. In: Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N. D., Weinberger, K. Q. (Eds.), Advances in Neural Information Processing Sy...

    2014

  15. [23]

    C., 2017

    Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A. C., 2017. Improved training of wasserstein gans. In: Guyon, I., Luxburg, U. V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (Eds.), Advances in Neural Information Processing Systems 30. ...

    2017

  16. [24]

    \'O ., 2010

    Hansson, G. \'O ., 2010. Consonant harmony: Long-distance interactions in phonology. University of California Press

    2010

  17. [25]

    Learning long-distance phonotactics

    Heinz, J., 2010. Learning long-distance phonotactics. Linguistic Inquiry 41 (4), 623--661. ://doi.org/10.1162/LING\_a\_00015

    2010 doi

  18. [26]

    Turkish vowel harmony

    Kabak, B., 2011. Turkish vowel harmony. In: van Oostendorp, M., Ewen, C. J., Hume, E., Rice, K. (Eds.), The Blackwell Companion to Phonology. Wiley Blackwell, Ch. 118, pp. 1--24. ://onlinelibrary.wiley.com/doi/abs/10.1002/9781444335262.wbctp0118

    2011 doi

  19. [27]

    Harmonic grammar: A formal multi-level connectionist theory of linguistic well-formedness: Theoretical foundations

    Legendre, G., Miyata, Y., Smolensky, P., 1990. Harmonic grammar: A formal multi-level connectionist theory of linguistic well-formedness: Theoretical foundations. University of Colorado, Boulder. ICS Technical Report \#90-5

    1990

  20. [28]

    MacMahon, M. K. C., 07 2013. Orthography and the early history of phonetics. In: Allan, K. (Ed.), The Oxford Handbook of the History of Linguistics. Oxford University Press, pp. 105--122. ://www.oxfordhandbooks.com/view/10.1093/oxfordhb/9780199585847.001.0001/oxfordhb-97801995...

    2013

  21. [29]

    F., 2001

    Marcus, G. F., 2001. The algebraic mind: Integrating connectionism and cognitive science. MIT press

    2001

  22. [30]

    F., Vijayan, S., Bandi Rao, S., Vishton, P

    Marcus, G. F., Vijayan, S., Bandi Rao, S., Vishton, P. M., 1999. Rule learning by seven-month-old infants. Science 283 (5398), 77--80. ://science.sciencemag.org/content/283/5398/77

    1999

  23. [31]

    L., Elman, J

    McClelland, J. L., Elman, J. L., 1986. The trace model of speech perception. Cognitive Psychology 18 (1), 1 -- 86. ://www.sciencedirect.com/science/article/pii/0010028586900150

    1986

  24. [32]

    L., Rumelhart, D

    McClelland, J. L., Rumelhart, D. E., , Group, P. R., 1986. Parallel distributed processing: Explorations in the microstructure of cognition. Vol. 2. MIT Press, Cambridge, MA

    1986

  25. [33]

    Inductive learning of locality relations in segmental phonology

    McMullin, K., Hansson, G., 2019. Inductive learning of locality relations in segmental phonology. Laboratory Phonology: Journal of the Association for Laboratory Phonology 10 (1), 14

    2019

  26. [34]

    C., Kello, C

    Plaut, D. C., Kello, C. T., 1999. The emergence of phonology from the interplay of speech comprehension and production: A distributed connectionist approach. In: The emergence of language. Lawrence Erlbaum Associates Publishers, Mahwah, NJ, US, pp. 381--415

    1999

  27. [35]

    Optimality Theory : Constraint Interaction in Generative Grammar

    Prince, A., Smolensky, P., 1993/2004. Optimality Theory : Constraint Interaction in Generative Grammar. Blackwell, Malden, MA, first published in 1993, Tech. Rep. 2, Rutgers University Center for Cognitive Science

    1993

  28. [36]

    R: A Language and Environment for Statistical Computing

    R Core Team , 2018. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria. ://www.R-project.org/

    2018

  29. [37]

    Unsupervised representation learning with deep convolutional generative adversarial networks

    Radford, A., Metz, L., Chintala, S., 2015. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434

    2015 arXiv

  30. [38]

    Analyzing distributional learning of phonemic categories in unsupervised deep neural networks

    R \"a s \"a nen, O., Nagamine, T., Mesgarani, N., 08 2016. Analyzing distributional learning of phonemic categories in unsupervised deep neural networks. CogSci ... Annual Conference of the Cognitive Science Society. Cognitive Science Society (U.S.). Conference 2016, 1757--176...

    2016

  31. [39]

    A typology of consonant agreement as correspondence

    Rose, S., Walker, R., 2004. A typology of consonant agreement as correspondence. Language 80 (3), 475--531. ://www.jstor.org/stable/4489721

    2004

  32. [40]

    E., McClelland, J

    Rumelhart, D. E., McClelland, J. L., Group, P. R., 1986. Parallel distributed processing: Explorations in the microstructure of cognition. Vol. 1. MIT Press, Cambridge, MA

    1986

  33. [41]

    Shain, C., Elsner, M., Jun. 2019. Measuring the perceptual availability of phonological features during language acquisition using unsupervised binary stochastic autoencoders. In: Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computat...

    2019

  34. [42]

    Regularization paths for cox's proportional hazards model via coordinate descent

    Simon, N., Friedman, J., Hastie, T., Tibshirani, R., 2011. Regularization paths for cox's proportional hazards model via coordinate descent. Journal of Statistical Software 39 (5), 1--13. ://www.jstatsoft.org/v39/i05/

    2011

  35. [43]

    Gradient symbolic representations in grammar: The case of french liaison

    Smolensky, P., Goldrick, M., 2016. Gradient symbolic representations in grammar: The case of french liaison. In: Rutgers Optimality Archive 1552, Rutgers University. pp. 1--37

    2016

  36. [44]

    Learning a gradient grammar of F rench liaison

    Smolensky, P., Rosen, E., Goldrick, M., 2019. Learning a gradient grammar of F rench liaison. In: Proceedings of the 2019 Annual Meeting on Phonology. pp. 1--12

    2019

  37. [45]

    Developing biases

    van de Vijver, R., Baer-Henney, D., 2014. Developing biases. Frontiers in Psychology 5, 634. ://www.frontiersin.org/article/10.3389/fpsyg.2014.00634

    2014

  38. [46]

    Discoverers of the phoneme

    van der Hulst, H., 07 2013. Discoverers of the phoneme. In: Allan, K. (Ed.), The Oxford Handbook of the History of Linguistics. Oxford University Press, pp. 167--191

    2013

  39. [47]

    van Schijndel, M., Mueller, A., Linzen, T., Nov. 2019. Quantity doesn ' t buy quality syntax with neural language models. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language P...

    2019

  40. [48]

    Preference for locality is affected by the prefix/suffix asymmetry

    White, J., Nevins, A., Polg \'a rdi, K., Martin, A., Kager, R., Linzen, T., Peperkamp, S., Topintzi, I., Markopoulos, G., van de Vijver, R., 2018. Preference for locality is affected by the prefix/suffix asymmetry. In: Hucklebridge, S., Nelson, M. (Eds.), NELS 48: Proceedings ...

    2018

  41. [49]

    N., 2011

    Wood, S. N., 2011. Fast stable restricted maximum likelihood and marginal likelihood estimation of semiparametric generalized linear models. Journal of the Royal Statistical Society (B) 73 (1), 3--36

    2011

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.