An Iterated Learning Model of language change that mixes supervised and unsupervised learning.
Jack Bunyan, Seth Bullock and Conor Houghton

Abstract

The iterated learning model is a simulation of how languages might evolve as they are passed from one generation to the next. It uses a highly simplified language: a mapping between vectors of zeros and ones representing meanings and signals. Nonetheless, it is possible to test whether the model language is compositional and expressive and, in the simulation, these properties arise though successive cycles of learning and generalization. However, while an agent's decoder map from signals to meanings is a neural network, the complementary encoder map is calculated in an unrealistic way. Here, we propose a more realistic model in which both the encoder and decoder are neural networks, concatenated together as an autoencoder, and pupils are required to learn from a mix of unsupervised and supervised examples, as children do. The model suggests that internal reflection on potential utterances is important in language learning and evolution.

Abstract as Gaeilge

Is é atá san iterated learning model ionsamhlú ar an bhealach a bhféadfadh teangacha éabhlóid a dhéanamh agus iad á n-aistriú ó ghlúin go glúin. Úsáideann sé teanga a simplíodh go mór: mapáil idir veicteoirí náideanna agus aonta a sheasann do bhríonna agus do chomharthaí. Mar sin féin, is féidir a thástáil an bhfuil an teanga samplach seo comhdhéanta eispriseach; san ionsamhlú, bíonn na hairíonna seo mar thoradh ar thimthriallta leantacha foghlama agus ginearálaithe. Fós féin, cé gur líonra néarach é léarscáil decoder an ghníomhaire ó chomharthaí go bríonna, ríomhtar an léarscáil chomhlántach encoder ar bhealach neamhréalaíoch. Molaimid samhail níos réalaíche anseo ina bhfuil an t-ionchódóir agus an díchódóir araon ina líonraí néaracha agus iad comhcheangailte le chéile mar autoencoder, agus is gá do dhaltaí foghlaim ó mheascán