Cleeremans, A. & McClelland, J.L. (1991). Learning the structure of event sequences. Journal of Experimental Psychology : General, 120, 235-253.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....one possible way to implement the drift equation (Eq. 3) 5 The essential feature is some form of recurrency, so that the previous state can 5 This is essentially the architecture used in the simple recurrent network of Elman (1990) and applied extensively to implicit serial learning tasks (Cleeremans McCleeland, 1991; Cleeremans, 1993) The di erence is in the speci cation of the weight matrices between layers. The simple recurrent networks include weight matrices modi able by back propagation between t IN and t i , between t i 1 and t i and between t i and an output layer. We have not yet discussed the ....

Cleeremans, A. & McCleeland, J. L. (1991). Learning the structure of event sequences. Journal of Experimental Psychology: General, 120, 235-253.

....hope that the SRN model could scale up. Conclusion This paper suggests that bilingual memory phenomena that have been explained by an interactive activation model of the BIA type (Grainger, 1993) may also be able to emerge from a simple recurrent connectionist network (SRN) model (Elman, 1990; Cleeremans McClelland, 1991; Cleeremans, 1993) The SRN model presented here receives as input a long, undifferentiated sequence of sentences in two micro languages and can reproduce a certain number of important effects that have been observed in studies of bilingual memory, in particular, those related to cross lingual ....

Cleeremans, A. & McClelland, J. (1991). Learning the structure of event sequences. Journal of Experimental Psychology: General, 120, 235-253.

....in turn was derived from its predecessor, context changes gradually. Figure 2 illustrates one possible way to implement the drift equation (Eq. 3) This is essentially the architecture used in the simple recurrent network of Elman (1990) and applied extensively to implicit serial learning tasks (Cleeremans McCleeland, 1991; Cleeremans, 1993) 6 The essential feature is some form of recurrency, so that the previous state can contribute to the new state. Although Figure 2 shows a separate set of units for the prior state and the current state, this long loop solution is not the only way to implement the drift ....

Cleeremans, A. & McCleeland, J. L. (1991). Learning the structure of event sequences. Journal of Experimental Psychology: General, 120, 235-253.

.... for the current input with respect to all the possible actions in an action centered module is done in a connectionist fashion through parallel spreading activation and thus highly ecient (such spreading of activation is assumed to be unconscious implicit by many, e.g. Hunt and Lansman 1986, Cleeremans and McClelland 1991, Bower 1996) We use a four layered connectionist network (see Figure 9) in which the rst three layers form a (either recurrent or feedforward) backpropagation network for computing Q values and the fourth layer (with only one node) performs stochastic decision making. The network is internally ....

A. Cleeremans and J. McClelland, (1991). Learning the structure of event sequences. Journal of Experimental Psychology: General. 120. 235-253.

....we allow the possibility of indirect verbalization of knowledge in the networks through some more elaborate transformation process. See section 3.5 for more analyses of this issue. 5 For arguments that more than distributed representation is needed for modeling explicit processes, see e.g. Cleeremans and McClelland (1991), although they themselves adopted distributed representation (for modeling implicit learning but not explicit learning) 2 DEVELOPING MODELS 9 edge can be learned in a variety of ways. This learning is di erent from the learning of procedural knowledge. Because of the representational di erence, ....

....of domains. The calculation of Q values for the current input with respect to all the possible actions is done in a connectionist fashion through parallel spreading activation and is thus highly ecient. Such spreading of activation is assumed to be implicit as, e.g. in Hunt and Lansman (1986) Cleeremans and McClelland (1991), and Bower (1996) We use a four layered connectionist network (see Figure 3) in which the rst three layers form a (either recurrent or feedforward) backpropagation network for computing Q values and the fourth layer (with only one node) performs stochastic decision making. The network is ....

[Article contains additional citation context not shown here]

A. Cleeremans and J. McClelland, (1991). Learning the structure of event sequences. Journal of Experimental Psychology: General. 120. 235-253.

.... for the current input with respect to all the possible actions in an action centered module is done in a connectionist fashion through parallel spreading activation and thus highly efficient (such spreading of activation is assumed to be unconscious implicit by many, e.g. Hunt and Lansman 1986, Cleeremans and McClelland 1991, Bower 1996) We use a four layered connectionist network (see Figure 9) in which the first three layers form a (either recurrent or feedforward) backpropagation network for computing Q values and the fourth layer (with only one node) performs stochastic decision making. The network is ....

A. Cleeremans and J. McClelland, (1991). Learning the structure of event sequences. Journal of Experimental Psychology: General. 120. 235-253.

....short term weights, although the same results would also hold in the latter case. Another appropriate domain involves short term memory and its consolidation into long term memory (e.g. Gardner Medwin, 1989) Goebel (1990) suggests how to use fast weights for serial rehearsal in short term memory. Cleeremans McClelland (1991) show how fast weights can account for the temporary biases of subjects in learning to respond to structured event sequences. This last work is particularly interesting because it involves specific biases towards recently occurring associations between stimuli, above and beyond the bias changes ....

Cleeremans, A. & McClelland, J. L. (1991). Learning the structure of event sequences. Journal of Experimental Psychology: General, 120:235--253.

....of future patterns are compared to T. A question arises as to how well a pattern should be learned. The rule of thumb used here is that the learning should halt when the target recall starts to deteriorate due to interference from the newly learnt pattern. The second hypothesis carries an echo of Cleeremans McClelland (1991) with respect to the rapid blurring of activations. Following each presentation of a pattern at the input layer of the network, the activation is not reset immediately, but gradually decays back to the neutral state. Successive patterns will then be affected by the temporally closest patterns. 4.1 ....

Cleeremans, A. and McClelland, J. (1991) Learning the structure of event sequences.

....better. A specific example of this complexity difference is as follows. Implicit learning of sequences (e.g. artificial grammar sequences) is biased towards sequences with a high level of statistical structure with much correlation (Stadler 1992) As has been demonstrated by Elman (1990) and by Cleeremans and McClelland (1991), recurrent backpropagation networks, as used in the bottom level of Clarion (in conjunction with Q learning) can handle sequences with complex statistical structures, given proper training procedures. Dienes (1992) reported similar results, in which a simple network model outperformed other ....

....recognition was achieved. In all of these cases, as suggested by Stanley et al. (1989) and Seger (1994) we may hypothesize that, due to the fact that explicit knowledge lags behind but improves along with implicit knowledge, explicit knowledge is in a way extracted from implicit knowledge. Cleeremans and McClelland (1991) also pointed out this possibility in discussing their data and models. Several developmental theorists have considered a similar process in child development. KarmiloffSmith (1986) suggested that developmental changes involve representational redescription. In young children, first low level ....

A. Cleeremans and J. McClelland, (1991). Learning the structure of event sequences. Journal of Experimental Psychology: General. 120. 235-253.

.... McLaren et al. 1989) Another appropriate domain involves short term memory and its consolidation into long term memory (e.g. Gardner Medwin, 1989) Goebel (1990) suggests how to use rapidly changing correlational weights for serial rehearsal in short term memory (see also Schmidhuber, 1992) Cleeremans McClelland (1991) show how short term weights can account for the temporary biases of subjects in learning to respond to structured event sequences. This last work is particularly interesting because it involves specific biases towards recently occurring associations between stimuli, above and beyond the bias ....

Cleeremans, A. & McClelland, J. L. (1991). Learning the structure of event sequences. Journal of Experimental Psychology: General, 120:235--253.

....little evidence of being aware that the material contained structure. Numerous subsequent studies of this effect have indicated that subjects can learn about complex sequential relationships despite remaining unable to fully elicit this knowledge in corresponding direct, explicit tasks (e.g. Cleeremans McClelland, 1991). One of the most convincing demonstrations of dissociations between conscious knowledge and behavior was obtained in a eye blink conditioning situation (Perruchet, 1985) In this experiment, people were exposed to a series of identical tones, 50 of which could be followed after a short interval ....

Cleeremans, A. & McClelland, J.L. (1991). Learning the structure of event sequences. Journal of Experimental Psychology : General, 120, 235-253.

....separately for valid (squares) invalid (circles) and neutral (triangles) cues, as well as for grammatical trials (filled symbols) and non grammatical trials (open symbols) the current context, and can thus be interpreted as representing preparation for the next event. Previous work (see Cleeremans McClelland, 1991; Cleeremans, 1993a, for detailed analysis of both processing in such networks and correspondence with human data) has shown that the SRN is able to account for about 80 of the variance in SRT data. To model performance in the experiments described above it is necessary to augment the SRN ....

....in turn connected with the output units. To assess how well this kind of network was able to account for SRT performance in this experiment, I conducted simulations in which the model was trained on the same material as human subjects and for the same number of trials, with the parameters used by Cleeremans and McClelland (1991). The network used local representations on both the input and output pools (i.e. each unit corresponded to one of the 6 stimuli or cues) To account for short term priming effects, the network used dual connection weights and running average activations on the output units, as described in ....

[Article contains additional citation context not shown here]

Cleeremans, A. & McClelland, J.L. (1991). Learning the structure of event sequences. Journal of Experimental Psychology: General, 120, 235--253.

.... of as involving elementary associative learning processes that result in a progressively developing sensitivity to the statistical constraints contained in the material (see also Stadler, 1992) Such processes are well instantiated by the Simple Recurrent Network (henceforth, SRN; see Elman, 1990; Cleeremans McClelland, 1991), which we describe later in this paper. In this context, Lee (1997) described an interesting sequence learning situation which, at first sight, seems to challenge traditional accounts of sequence learning. Indeed, the stimulus set used by Lee consisted of a random selection of the 720 (6 ) ....

....come to approximate the optimal conditional probabilities associated with their appearance in the current context, and can thus be interpreted as representing implicit preparation for the next element when the network is used as a model of human sequence learning performance. Previous work (see Cleeremans McClelland, 1991; Cleeremans, 1993) has shown that the SRN is able to account for about 80 of the variance in sequential choice reaction time data. Simulation parameters and procedure To assess how well the SRN could capture RT performance in this experiment, we trained the model on the same material as human ....

[Article contains additional citation context not shown here]

Cleeremans, A. & McClelland, J.L. (1991). Learning the structure of event sequences. Journal of Experimental Psychology: General, 120, 235-253.

....manner without encoding rules explicitly. Likewise, the performance of symbolic computational systems based on chunking (Servan Schreiber Anderson, 1990; Laird, Rosenbloom, Newell, 1985; Rosenbloom, Newell, Laird, 1990) overlaps largely with the performance of the Simple Recurrent Network (Cleeremans McClelland, 1991) in accounting for artificial grammar learning tasks performance (see Berry and Dienes, 1993; Cleeremans, 1993 for discussions) Some authors even go as far as claiming that many of these models are not empirically differentiable (Barsalou, 1990; Goldstone Krushke, 1994) In contrast to these ....

....neurons ) The obvious advantage of this interactive method is that it is much easier to explore the space of possible solutions in a computer model than in real humans. Finally, a somewhat more detailed example illustrates how modeling work was actually instrumental in understanding the data. Cleeremans McClelland (1991) explored performance in a reaction time situation characterized by the fact that the locations at which successive events appeared were determined based on the generation rules specified by a probabilistic finite state automaton. Thus, on each trial, the stimulus could appear at any screen ....

Cleeremans, A. & McClelland, J. (1991). Learning the structure of event sequences. Journal of Experimental Psychology: General. 120, 235--253.

....memory until processing of the embedding information has been completed. Such expressions present interesting challenges for popular sequential connectionist architectures such as the Simple Recurrent Network (henceforth, SRN) The SRN, first proposed by Elman (1990) and subsequently adapted by Cleeremans McClelland (1991) to simulate sequential effects in reaction time tasks, is shown in Figure 1. The network uses back propagation to learn to predict the next element of a sequence based only on the current element and on a representation of the temporal context that the network has elaborated itself. To do so, it ....

....come 2 to approximate the optimal conditional probabilities associated with their appearance in the current context, and can thus be interpreted as representing implicit preparation for the next event when the network is used as a model of human sequence learning performance. Previous work (see Cleeremans McClelland, 1991; Cleeremans, 1993) has shown that the SRN is able to account for about 80 of the variance in sequential choice reaction time data. The SRN, however, also suffers from an important limitation in its ability to learn sequential material. Indeed, one key aspect of learning in the SRN is that the ....

[Article contains additional citation context not shown here]

Cleeremans, A. & McClelland, J.L. (1991). Learning the structure of event sequences. JEP:G, 120, 235-253.

....multiply earnings during the SRT task. This parameter was set to be 1.0 plus the proportion by which accuracy exceeded chance level during Cleeremans: Sequence learning with multiple cues 10 generation. Stimulus and Cue Generation. Stimulus generation followed the general design described in Cleeremans and McClelland (1991) and in Jim nez, Mend z and Cleeremans (in press) Stimuli were generated based on the noisy finite state grammar used by Jim nez, Mend z Cleeremans (in press) and illustrated in Figure 1. Finite state grammars consist of nodes linked by labeled arcs. A sequence of labels can be generated by ....

....any stimulus to be involved in a direct repetition of itself. This guarantees that RT effects are not contaminated by short term priming effects, which have large facilitatory effects on performance that are completely independent from the factors of interest in this research (see Bertelson, 1961; Cleeremans McClelland, 1991; Hyman, 1953) Results Participants were exposed to 10 sessions of an RT task and were subsequently asked to try to predict each successive event in a generation task. I first present the RT data. Cleeremans: Sequence learning with multiple cues 12 Reaction Time task performance Figure 2 ....

[Article contains additional citation context not shown here]

Cleeremans, A. & McClelland, J.L. (1991). Learning the structure of event sequences. Journal of Experimental Psychology: General, 120, 235-253.

.... and some authors go as far as claiming that they are not empirically differentiable (Barsalou, 1990; Goldstone Krushke, 1994) The performance of symbolic systems based on chunking (Servan Schreiber Anderson, 1990) overlaps largely with the performance of the Simple Recurrent Network (Cleeremans McClelland, 1991) in artificial grammar learning tasks (see Berry and Dienes, 1993) Dienes (in press) also compared Logan s (1988) instance based model with a reinforcement based connectionist model (Barto, Sutton Anderson, 1983) in the context of process control tasks, and again found a large overlap in how ....

....but merely give an example of their application taken from research by Jim nez et al. in press) Jim nez et al. in press) explored the relationship between reaction time performance and explicit knowledge as revealed through a subsequent generation task. The reaction time task was similar to Cleeremans and McClelland s (1991) situation in involving sequential material generated based on a probabilistic finite state grammar. Through detailed partial correlational analyses (which controlled for knowledge expressed through the generation task) of the relationship between performance at the reaction time task and the ....

[Article contains additional citation context not shown here]

Cleeremans, A. & McClelland, J.L. (1991). Learning the structure of event sequences. Journal of Experimental Psychology: General 120: 235--253.

....found at a global level of analysis, dissociations can still be obtained when more detailed analyses are conducted. Finally, we also show that subjects are limited in the depth of the contingencies they can learn about, and that these limitations are shared by the Simple Recurrent Network model of Cleeremans McClelland (1991). Introduction Over the last few years, three different paradigms have become standard in the field of implicit learning. Each has produced a large body of evidence suggesting that subjects can develop sensitivity to complex stimulus covariations without intention to learn or without awareness ....

....implicit or explicit learning processes The goal of this paper is to reflect upon the conditions required to demonstrate unconscious learning and to present new experimental data that aim to fulfill these conditions. We also present simulation work using the Simple Recurrent Network (Elman, 1990; Cleeremans McClelland, 1991, hereafter C McC) and discuss which challenges these data pose for the model. Many authors have explored the theoretical and methodological flaws underlying the widespread assumption that some given measure of performance may be taken as an absolute index of awareness (e.g. Jacoby, 1991; ....

Cleeremans, A. & McClelland, J.L. (1991). Learning the structure of event sequences. JEP:G 120, 235-253.

....inspection, and that are differentially affected by the availability of attentional resources. In this paper, I propose a new information processing model of sequence learning and explore how well it can account for these data. The model is based on the Simple Recurrent Network (Elman, 1990; Cleeremans McClelland, 1991; Cleeremans, 1993) which it extends by allowing additional information to modulate processing. The model implements the notion that awareness of sequence structure changes the task from one of anticipating the next event based on temporal context to one of retrieving the next event from ....

....representations become less reliable, as when a secondary task is performed concurrently with the sequence learning task. Introduction In recent years, sequence learning in choice reaction settings has elicited considerable interest as a vehicle to study implicit information processing (e.g. Cleeremans McClelland, 1991; Lewicki, Hill, Bizot, 1988; Nissen and Bullemer, 1987; Perruchet, Amorim, 1992) In such tasks, subjects are presented with a visuo spatial choice reaction task, but, unknown to them, the sequence of successive stimuli is structured, so that the uncertainty about the next event may be ....

[Article contains additional citation context not shown here]

Cleeremans, A. & McClelland, J.L. (1991). Learning the structure of event sequences. Journal of Experimental Psychology: General, 120, 235-253.

....this knowledge is exclusively responsible for performance, then it becomes indeed hard to understand what is implicit about implicit learning. But I find it rather implausible that subject s knowledge can be adequately accounted for by such theories. Consider for instance a subject performing in Cleeremans McClelland s (1991) experiments. In sharp contrast to the simple repeating short deterministic sequences used in the vast majority of sequence learning experiments, the stimulus material we used was generated from a non deterministic finite state grammar. Hence, almost all permutations between elements of ....

....with each cluster representing a node of the grammar as abstract a representation as could be. In other cases, however, the network s internal representations tend to be organized in numerous very small clusters that each correspond to one or to a few training instances (see Servan Schreiber, Cleeremans and McClelland, 1991, for detailed examples) As a case in point, the SRN has often been described as processing fragmentary information. This is descriptively correct, but it is not how things work inside the network. The network does not develop a database of subsequences that it can consult and ponder about as a ....

[Article contains additional citation context not shown here]

Cleeremans, A. & McClelland, J.L. (1991). Learning the structure of event sequences. Journal of Experimental Psychology: General 120: 235--253.

.... (Barsalou, 1990; Goldstone Krushke, 1994) Likewise, the performance of symbolic computational systems based on chunking (Servan Schreiber Anderson, 1990; Laird, Rosenbloom, Newell, 1985; Rosenbloom, Newell, Laird, 1990) overlaps largely with the performance of the Simple Recurrent Network (Cleeremans McClelland, 1991) in accounting for artificial grammar CLEEREMANS FRENCH 7 learning tasks performance (see Berry Dienes, 1993; Cleeremans, 1993, for discussions) What are we to make of these overlapping models Should some be taken as wrong and others as correct, even though they are all equally successful ....

....real neurons ) The obvious advantage of this interactive method is that it is much easier to explore the space of possible solutions in a computer model than in real humans. Finally, here is a somewhat more detailed example of how modeling work was actually instrumental in understanding the data. Cleeremans McClelland (1991) explored performance in a reaction time situation characterized by the fact that the locations at which successive events appeared were determined based on the generation rules specified by a probabilistic finite state automaton. Thus on each trial, the stimulus could appear at any screen ....

Cleeremans, A. & McClelland, J.L. (1991). Learning the structure of event sequences. Journal of Experimental Psychology: General 120: 235-- 253.

No context found.

A. Cleeremans and J. McClelland, (1991). Learning the structure of event sequences. Journal of Experimental Psychology: General. 120. 235-253.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC