<?xml version="1.0" standalone="yes"?>
<Paper uid="W97-1009">
  <Title>Evolution of a Rapidly Learned Representation for Speech</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
5 Discussion
</SectionTitle>
    <Paragraph position="0"> By developing an appropriate architecture, timeconstants and learning rules over many generations, the task of learning to represent speech sounds is made more rapid over the course of development of an individual network. Evolution does all the hard work and gives the network a developmental &amp;quot;legup&amp;quot;. However, having the correct innate architecture and learning rules is not sufficient for creating good representations. Weights are not inherited between generations so the network is dependent on the environment for learning the correct representation. If deprived of sound input or fed acoustically filtered speech input, the model cannot form meaningful representations because each network starts life with a random set of weights. But given the sort of auditory input heard by an infant the model rapidly creates the same set of universal features, whether or not it is in a noisy environment and whatever the language it hears.</Paragraph>
    <Paragraph position="1"> We envisage that this method of creating a quick and dirty initial representation of sounds by innately guided learning is not specific to humans. Clearly, humans and other animals have not been selected for their ability to discriminate the phonemes of English. But we would expect results similar to those presented here if the selection criterion were the ability to discriminate a wide range of spectrally dissimilar sounds in the environment from only limited exposure to their patterns of regularity e.g. discrimination of the maternal call from other conspecific calls, and the sound of predators from everyday environmental noises. It is therefore unsurprising that animals have been found, after suitable training, to discriminate some phonemes in similar ways as do humans (Kuhl &amp; Miller, 1975).</Paragraph>
    <Paragraph position="2"> The advantages of innately guided learning over other self-organising networks are that it is much faster and is less dependent on the &amp;quot;correct&amp;quot; environmental statistics. It also offers an account of how infants from different linguistic environments can come up with the same featural representation so soon after birth. In this sense innately guided learning as implemented in this model shows how genes and the environment could interact to ensure rapid development of a featural representation of speech on which further linguistic development depends.</Paragraph>
  </Section>
class="xml-element"></Paper>