Tuesdays & Thursdays, 2-3:20pm in SSL 206
Instructors: Lisa Pearl, Department of Cognitive Sciences, SBSG 2314.
Mark Steyvers, Department of Cognitive Sciences, SBSG 2316.
Office Hours for Lisa: Wednesday 1:30pm - 3:00pm, and by appointment.
Email is the best way to reach her to schedule an appointment
not during regular office hours.
Language is an amazingly complex system of knowledge that must be learned from noisy input. That very young children who don't have the cognitive sophistication to count to four can accomplish this task is no small wonder. That adults who have far more cognitive resources and learning strategies - not to mention knowledge of their native language - can't is perhaps even more striking. How is this possible?
Some researchers, primarily linguists, propose that children must bring innate biases to guide their learning, and these biases are later lost during maturation. These may be biases that restrict what hypotheses the child considers, for instance, or biases that cause the child to heed certain kinds of data over others. Other researchers have proposed that many language learning problems can be solved without prior language-specific knowledge. They would argue that general learning mechanisms humans may have available (often originating from machine learning and artificial intelligence) are more powerful than was previously thought possible, removing the need for domain-specific biases to guide learning.
In this class, we explore computational models of language learning that allow both sides to speak to each other. We will examine learning models from psychology, machine learning, and artificial intelligence, and see how they fare on some of the case studies that first caused researchers to posit the necessity of innate biases for language learning. How far can these techniques take us? Where do innate biases seem necessary, and where can general learning mechanisms accomplish what humans do in just the way that humans do it? In all cases, we ground the learning models in available empirical data and consider the psychological plausibility of the learning models – for in the end, we want to understand how it is that the human mind accomplishes this task, and in particular, the mind of a very young child.
We will be reading a number of research articles. A bibliography of these articles can be found on the readings section, and accessed through the schedule page (provided you have the class username and password).
Announcements:
Note: All readings can be accessed using the username and password received in the first class session. Of course, you can also always track down these articles yourself in most cases. Look to the schedule, and be thinking about what papers/topics you'd like to present to the class.