Raquel G. Alhama

Publications

Displaying 1 - 11 of 11
  • Alhama, R. G., & Zuidema, W. (2019). A review of computational models of basic rule learning: The neural-symbolic debate and beyond. Psychonomic Bulletin & Review, 26(4), 1174-1194. doi:10.3758/s13423-019-01602-z.

    Abstract

    We present a critical review of computational models of generalization of simple grammar-like rules, such as ABA and ABB. In particular, we focus on models attempting to account for the empirical results of Marcus et al. (Science, 283(5398), 77–80 1999). In that study, evidence is reported of generalization behavior by 7-month-old infants, using an Artificial Language Learning paradigm. The authors fail to replicate this behavior in neural network simulations, and claim that this failure reveals inherent limitations of a whole class of neural networks: those that do not incorporate symbolic operations. A great number of computational models were proposed in follow-up studies, fuelling a heated debate about what is required for a model to generalize. Twenty years later, this debate is still not settled. In this paper, we review a large number of the proposed models. We present a critical analysis of those models, in terms of how they contribute to answer the most relevant questions raised by the experiment. After identifying which aspects require further research, we propose a list of desiderata for advancing our understanding on generalization.
  • Alhama, R. G., Siegelman, N., Frost, R., & Armstrong, B. C. (2019). The role of information in visual word recognition: A perceptually-constrained connectionist account. In A. Goel, C. Seifert, & C. Freksa (Eds.), Proceedings of the 41st Annual Meeting of the Cognitive Science Society (CogSci 2019) (pp. 83-89). Austin, TX: Cognitive Science Society.

    Abstract

    Proficient readers typically fixate near the center of a word, with a slight bias towards word onset. We explore a novel account of this phenomenon based on combining information-theory with visual perceptual constraints in a connectionist model of visual word recognition. This account posits that the amount of information-content available for word identification varies across fixation locations and across languages, thereby explaining the overall fixation location bias in different languages, making the novel prediction that certain words are more readily identified when fixating at an atypical fixation location, and predicting specific cross-linguistic differences. We tested these predictions across several simulations in English and Hebrew, and in a pilot behavioral experiment. Results confirmed that the bias to fixate closer to word onset aligns with maximizing information in the visual signal, that some words are more readily identified at atypical fixation locations, and that these effects vary to some degree across languages.
  • Zuidema, W., French, R. M., Alhama, R. G., Ellis, K., O'Donnell, T. J. O., Sainburgh, T., & Gentner, T. Q. (2019). Five ways in which computational modeling can help advance cognitive science: Lessons from artificial grammar learning. Topics in Cognitive Science. Advance online publication. doi:10.1111/tops.12474.

    Abstract

    There is a rich tradition of building computational models in cognitive science, but modeling, theoretical, and experimental research are not as tightly integrated as they could be. In this paper, we show that computational techniques—even simple ones that are straightforward to use—can greatly facilitate designing, implementing, and analyzing experiments, and generally help lift research to a new level. We focus on the domain of artificial grammar learning, and we give five concrete examples in this domain for (a) formalizing and clarifying theories, (b) generating stimuli, (c) visualization, (d) model selection, and (e) exploring the hypothesis space.
  • Alhama, R. G., & Zuidema, W. (2018). Pre-Wiring and Pre-Training: What Does a Neural Network Need to Learn Truly General Identity Rules? Journal of Artificial Intelligence Research, 61, 927-946. doi:10.1613/jair.1.11197.

    Abstract

    In an influential paper (“Rule Learning by Seven-Month-Old Infants”), Marcus, Vijayan, Rao and Vishton claimed that connectionist models cannot account for human success at learning tasks that involved generalization of abstract knowledge such as grammatical rules. This claim triggered a heated debate, centered mostly around variants of the Simple Recurrent Network model. In our work, we revisit this unresolved debate and analyze the underlying issues from a different perspective. We argue that, in order to simulate human-like learning of grammatical rules, a neural network model should not be used as a tabula rasa, but rather, the initial wiring of the neural connections and the experience acquired prior to the actual task should be incorporated into the model. We present two methods that aim to provide such initial state: a manipulation of the initial connections of the network in a cognitively plausible manner (concretely, by implementing a “delay-line” memory), and a pre-training algorithm that incrementally challenges the network with novel stimuli. We implement such techniques in an Echo State Network (ESN), and we show that only when combining both techniques the ESN is able to learn truly general identity rules. Finally, we discuss the relation between these cognitively motivated techniques and recent advances in Deep Learning.
  • Alhama, R. G., & Zuidema, W. (2017). Segmentation as Retention and Recognition: the R&R model. In G. Gunzelmann, A. Howes, T. Tenbrink, & E. Davelaar (Eds.), Proceedings of the 39th Annual Conference of the Cognitive Science Society (CogSci 2017) (pp. 1531-1536). Austin, TX: Cognitive Science Society.

    Abstract

    We present the Retention and Recognition model (R&R), a probabilistic exemplar model that accounts for segmentation in Artificial Language Learning experiments. We show that R&R provides an excellent fit to human responses in three segmentation experiments with adults (Frank et al., 2010), outperforming existing models. Additionally, we analyze the results of the simulations and propose alternative explanations for the experimental findings.
  • Stanojevic, M., & Alhama, R. G. (2017). Neural discontinuous constituency parsing. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing (pp. 1666-1676). Association for Computational Linguistics.

    Abstract

    One of the most pressing issues in dis- continuous constituency transition-based parsing is that the relevant information for parsing decisions could be located in any part of the stack or the buffer. In this pa- per, we propose a solution to this prob- lem by replacing the structured percep- tron model with a recursive neural model that computes a global representation of the configuration, therefore allowing even the most remote parts of the configura- tion to influence the parsing decisions. We also provide a detailed analysis of how this representation should be built out of sub-representations of its core elements (words, trees and stack). Additionally, we investigate how different types of swap or- acles influence the results. Our model is the first neural discontinuous constituency parser, and it outperforms all the previ- ously published models on three out of four datasets while on the fourth it obtains second place by a tiny difference.

    Supplementary material

    http://aclweb.org/anthology/D17-1174
  • Alhama, R. G., & Zuidema, W. (2016). Pre-Wiring and Pre-Training: What does a neural network need to learn truly general identity rules? In T. R. Besold, A. Bordes, & A. D'Avila Garcez (Eds.), CoCo 2016 Cognitive Computation: Proceedings of the Workshop on Cognitive Computation: Integrating neural and symbolic approaches 2016. CEUR Workshop Proceedings.

    Abstract

    In an influential paper, Marcus et al. [1999] claimed that connectionist models cannot account for human success at learning tasks that involved generalization of abstract knowledge such as grammatical rules. This claim triggered a heated debate, centered mostly around variants of the Simple Recurrent Network model [Elman, 1990]. In our work, we revisit this unresolved debate and analyze the underlying issues from a different perspective. We argue that, in order to simulate human-like learning of grammatical rules, a neural network model should not be used as a tabula rasa , but rather, the initial wiring of the neural connections and the experience acquired prior to the actual task should be incorporated into the model. We present two methods that aim to provide such initial state: a manipu- lation of the initial connections of the network in a cognitively plausible manner (concretely, by implementing a “delay-line” memory), and a pre-training algorithm that incrementally challenges the network with novel stimuli. We implement such techniques in an Echo State Network [Jaeger, 2001], and we show that only when combining both techniques the ESN is able to learn truly general identity rules.
  • Alhama, R. G., & Zuidema, W. (2016). Generalization in Artificial Language Learning: Modelling the Propensity to Generalize. In Proceedings of the 7th Workshop on Cognitive Aspects of Computational Language Learning (pp. 64-72). Association for Computational Linguistics. doi:10.18653/v1/W16-1909.

    Abstract

    Experiments in Artificial Language Learn- ing have revealed much about the cogni- tive mechanisms underlying sequence and language learning in human adults, in in- fants and in non-human animals. This pa- per focuses on their ability to generalize to novel grammatical instances (i.e., in- stances consistent with a familiarization pattern). Notably, the propensity to gen- eralize appears to be negatively correlated with the amount of exposure to the artifi- cial language, a fact that has been claimed to be contrary to the predictions of statis- tical models (Pe ̃ na et al. (2002); Endress and Bonatti (2007)). In this paper, we pro- pose to model generalization as a three- step process, and we demonstrate that the use of statistical models for the first two steps, contrary to widespread intuitions in the ALL-field, can explain the observed decrease of the propensity to generalize with exposure time.
  • Alhama, R. G., Scha, R., & Zudema, W. (2015). How should we evaluate models of segmentation in artificial language learning? In N. A. Taatgen, M. K. van Vugt, J. P. Borst, & K. Mehlhorn (Eds.), Proceedings of ICCM 2015 (pp. 172-173). Groningen: University of Groningen.

    Abstract

    One of the challenges that infants have to solve when learn- ing their native language is to identify the words in a con- tinuous speech stream. Some of the experiments in Artificial Grammar Learning (Saffran, Newport, and Aslin (1996); Saf- fran, Aslin, and Newport (1996); Aslin, Saffran, and Newport (1998) and many more) investigate this ability. In these ex- periments, subjects are exposed to an artificial speech stream that contains certain regularities. Adult participants are typ- ically tested with 2-alternative Forced Choice Tests (2AFC) in which they have to choose between a word and another sequence (typically a partword, a sequence resulting from misplacing boundaries).
  • Alhama, R. G., Scha, R., & Zuidema, W. (2014). Rule learning in humans and animals. In E. A. Cartmill, S. Roberts, H. Lyn, & H. Cornish (Eds.), The evolution of language: Proceedings of the 10th International Conference (EVOLANG 10) (pp. 371-372). Singapore: World Scientific.
  • Marti, M., Alhama, R. G., & Recasens, M. (2012). Los avances tecnológicos y la ciencia del lenguaje. In T. Jiménez Juliá, B. López Meirama, V. Vázquez Rozas, & A. Veiga (Eds.), Cum corde et in nova grammatica. Estudios ofrecidos a Guillermo Rojo (pp. 543-553). Santiago de Compostela: Universidade de Santiago de Compostela.

    Abstract

    La ciencia moderna nace de la conjunción entre postulados teóricos y el desarrollo de una infraestructura tecnológica que permite observar los hechos de manera adecuada, realizar experimentos y verificar las hipótesis. Desde Galileo, ciencia y tecnología han avanzado conjuntamente. En el mundo occidental, la ciencia ha evolucionado desde pro-puestas puramente especulativas (basadas en postulados apriorísticos) hasta el uso de métodos experimentales y estadísticos para explicar mejor nuestras observaciones. La tecnología se hermana con la ciencia facilitando al investigador una aproximación adecuada a los hechos que pretende explicar. Así, Galileo, para observar los cuerpos celestes, mejoró el utillaje óptico, lo que le permitió un acercamiento más preciso al objeto de estudio y, en consecuencia, unos fundamentos más sólidos para su propuesta teórica. De modo similar, actualmente el desarrollo tecnológico digital ha posibilitado la extracción masiva de datos y el análisis estadístico de éstos para verificar las hipótesis de partida: la lingüística no ha podido dar el paso desde la pura especulación hacia el análisis estadístico de los hechos hasta la aparición de las tecnologías digitales.

Share this page