A Critical Look at the Mechanisms Underlying Implicit Sequence Learning Todd M Gureckis gureckis love psy utexas edu Bradley C Love love psy utexas edu Department of Psychology The University of Texas at Austin Austin TX 78712 USA Abstract In this report a model of human sequence learning is developed called the linear associative shift register LASR LASR uses a simple error driven associative learning rule to incrementally acquire information about the structure of event sequences In contrast to recent modeling approaches to implicit sequence learning LASR describes learning as a simple and limited process We argue that this simplicity is a virtue in that the complexity of the model is better matched to the demonstrated complexity of human processing The model is applied in a variety of situations including implicit learning via the serial reaction time SRT task and statistical word learning The results of these simulations highlight commonalities between different tasks and learning modalities which suggest similar underlying learning mechanisms Introduction One of the most striking aspects of human behavior is the ease with which we can acquire new skills with little conscious effort In order to better understand this phenomena a large literature has developed exploring the ability of participants to implicitly learn about the sequential structure of a series of events see Cleeremans Destrebecqz Boyer 1998 for a review However the type of memory and learning mechanisms which might support such learning are not well understood see Keele Ivry Mayr Hazeltine Heuer 2004 or Sun Sluzarz Terry 2005 for some recent proposals In this paper we develop a simple model of sequence learning behavior called the linear associative shift register LASR The model is unique from past approaches in that it describes implicit sequence learning as a simple and limited process which operates on a small temporary buffer of past events This contrasts with other models of sequence learning which have described learning as a more complex and flexible process Cleeremans McClelland 1991 Cleeremans 1993 Sun et al 2005 There are two main goals of this report First we explore the ability of this simple model to account for sequential learning phenomena in a variety of implicit learning situations including the serial reaction time SRT task and statistical word learning paradigms LASR provides a similar account of the type of processing which underlies performance in both kinds of tasks suggesting that they may rely on similar underlying mechanisms Second we demonstrate how a very simple learning mechanism such as LASR can provide a detailed account of a number of findings from the implicit sequence learning literature A key criticism we develop is that in previous modeling accounts such as the simple recurrent network SRN of Cleeremans 1993 the complexity of the model is not well matched to the demonstrated complexity of the learner While LASR cannot explain all aspects of our rich sequential behavior we believe the model provides a unique baseline against which to test more complex theories and experiments We begin by introducing the LASR model and the principles upon which it is based Next we consider a study conducted by Lee 1997 assessing implicit learning of sequentially structured material Finally we explore the ability of LASR to account for statistical word learning in infants as reported by Saffran Aslin and Newport 1996 The Linear Associative Shift Register LASR Model LASR is a mechanistic model of implicit sequence learning The model describes implicit sequence learning as the task of appreciating the associative relationship between past events and future ones LASR assumes that subjects maintain a limited memory for the sequential order of past events and that they use a simple error driven associative learning rule Widrow Hoff 1960 Rescorla Wagner 1972 to incrementally acquire information about sequential structure Despite it s simplicity the model can very quickly learn to appreciate rather complex dependencies between events which are structured in time The model is organized around 3 principles 1 Past events are stored in a temporary buffer The model begins by assuming a simple shift register memory for past events Individual elements of the register are referred to as slots New events encountered in time are inserted at one end of the register and all past events are accordingly shifted one time slot Thus the most recent event is always located in the right most slot of the register see Figure 1 This form of memory maintains the sequential order of recent events using spatial position see Sejnowski and Rosenberg 1987 or Cleereman s 1993 buffer network for similar approaches 2 Learning to predict what comes next This simple memory mechanism forms the basis of a detector see Figure 1 A detector is a simple single layer linear network or perceptron Rosenblatt 1958 which learns to predict the occurrence of a single future event based on past events Because each detector predicts only a single event a separate detector is needed for each possible event Each detector has a weight from each event outcome at each time slot On each trial activation from each memory register slot is passed over a connection weight and summed to compute the activation of the detector s prediction unit The task of a detector is to adjust the weights from individual memory slots so that it can successfully predict the future occurrence of it s response Each detector learns to strengthen the connection weights for memory slots which prove predictive of the detector s response while weakening those which are not predictive or are counter predictive 2 Recent events have more influence on learning than past events The model assumes that events in the recent past are remembered better than events which happened long ago This effect is implemented by attenuating the activation strength of each register position by how far back in time the event occurred Because of this an event which happened at time t 1 has more influence on future predictions than events which happened at t 2 t 3 etc Similarly learning is slower for slots which are positioned further in the past because their activation strength is reduced see Equation 3 Model Formalism The following section describes the mathematical formalism of the model The model is easily described using three equations and three intuitive parameters Memory As illustrated at the top of
View Full Document
Unlocking...