How do humans learn, and how can an
evolving new science of learning shape the way humans and machines are
taught in the future?
Jeffrey Heinz, a University of Delaware assistant professor of linguistics and cognitive science, explores these issues as co-author of Sentence and Word Complexity, an article that appears in the July 15, 2011, issue of Science magazine, the journal published by the American Association for the Advancement of Science.
William Idsardi, associate professor in the Department of Linguistics
and Program in Neuroscience and Cognitive Science at the University of
Maryland College Park and former chair of what was then the UD
Department of Linguistics, is also co-author of the article.
Heinz noted that this new science of learning reflects a collective
effort by psychologists, neuroscientists, linguists, computer
scientists, roboticists, and others, to understand how humans learn and
how to make machines learn.
The new part in the learning problem comes from new techniques that
have cropped up in the last few decades, Heinz said. Its a very
exciting area, whose results are bound to shape the daily lives of
future generations.
The methodologies employed range from experimentation on human
adults, children and infants, to computational modeling and mathematical
proofs, Heinz said.
The article asks if there are differences among learning patterns
during language perception, and if so, are these differences best
explained by specialized or non-specialized learning mechanisms.
We can measure the complexity of sound patterns, and researchers
have found that the sound patterns are simpler than syntax patterns,
Heinz said. It may mean, as we suggest in the article, that these kinds
of patterns are learned in different kinds of ways than sentence
patterns.
Natural human language, the authors note, distinguishes between well-formed and ill-formed sentences and words.
When most people think of a languages grammar, they think of rules
that determine how words go together, such as in Spanish, when we say
el gato blanco, for the white cat, or la princessa blanca, for
the white princess, Heinz said. There also are rules that determine
how words can sound in a language, which is why English speakers can
coin words like bling, but not gding.
The article suggests the possibility that the properties of sentence
patterns and sound patterns reflect the properties of the ways in which
humans learn languages.
When we learn something, we make a leap from our observations to
some general pattern or rule, Heinz said. You have to leap somewhere.
The idea is, when learning phonology, we leap in one way, but when
learning syntax, we leap in another.
While research continues, work in psychology and linguistics has
determined that infants have picked up the key characteristics of sound
patterns during the first year of life, before they can talk, Heinz
said.
Its hard to know how much syntax they have picked up in this time,
Heinz said. Generally, linguists and psychologists find that the
children have internalized the rules governing the sound and sentence
patterns of their language by around five years of age.
The possibility that distinct learning mechanisms are responsible for
learning sentence patterns and sound patterns indicates a certain kind
of division of labor is occurring in the human brain, Heinz said.
Such findings might cause machine designers to look for a specialized
learning algorithm instead of a general purpose one, Heinz said.
In the same way that humans dont have a general sensory organ that
solves the problem of sensing; we have ears for earing and eyes for
seeing, etc., Heinz said. We also dont have a general learning
strategy. Instead, we have multiple learning strategies, and,
unconsciously, we use different ones on different kinds of data.