You are here: Home Events MPI Colloquium series - Folder MPI Colloquium Series - 2017 Emmanuel Dupoux, September 5

Emmanuel Dupoux, September 5

Reverse Engineering Infant Language Acquisition using Machine Learning

Abstract

The mechanisms that enable infants to spontaneously acquire their native(s) language(s) are still not well understood. I examine the conditions under which recent progress in machine learning can provide insights on these mechanisms. I first argue that computational models of language acquisition must address the full complexity of language learning, ie. address the variability and ambiguity of raw sensory signals, instead of dealing with artificial or simplified language inputs. I then show that although deep neural networks can overcome the variability of raw sensory signals, they rely on implausible supervised learning mechanisms. I  summarize recent work in the unsupervised discovery of linguistic units from continuous speech (phonemes and words), and the use of these algorithms to test hypotheses about the functional role of infant-directed speech.

 

 Date:                     Tuesday, September 5, 2017

Time:                     15h45

Location:              MPI Conference Room 163

                               Wundtlaan 1

                               6525 XD Nijmegen

 

Last checked 2017-09-04 by edith sjoerdsma

Street address
Wundtlaan 1
6525 XD Nijmegen
The Netherlands


Mailing address
P.O. Box 310
6500 AH Nijmegen
The Netherlands

Phone:   +31-24-3521911
Fax:        +31-24-3521213
E-mail:   


Public Outreach Officer
Charlotte Horn