signature=71bf3c90ff11a93dbd596310980ef7e4,Discovering Processing Stages by combining EEG with Hidde...

摘要：

Discovering Processing Stages by combining EEG with Hidden Markov Models Jelmer P. Borst (jelmer@cmu.edu) John R. Anderson (ja+@cmu.edu) Dept. of Psychology, Carnegie Mellon University Abstract with condition. Using this information, and by comparing EEG signatures between states and experimental conditions, one can interpret the functional characteristics of the identified processing stages. Our approach is based on a similar method that was used to analyze fMRI data (Anderson & Fincham, in press; Anderson et al., 2010). For instance, Anderson and Fincham (in press) applied the method to mathematical problem solving, and discovered four stages: encoding the problems, planning a solution strategy, solving the problems, and entering a response. Although these results were promising, the temporal resolution of fMRI is severely limited, both by having scans that typically last one to two seconds and by the sluggish nature of the hemodynamic response. EEG, on the other hand, has a millisecond resolution, allowing for the discovery of processing stages in fast-paced tasks. We applied the HMM-EEG analysis to an associative recognition task. During the study phase of this task, subjects were asked to learn word pairs. In a subsequent test phase – during which EEG data were collected – subjects were again presented with word pairs, which could be the same pairs as they learned previously (targets), rearranged pairs (re-paired foils), or pairs consisting of novel words (new foils). Subjects had to decide whether they had seen the pair during the study phase or not. Successful discrimination required remembering not only that the words were studied (item information), but also how the words were paired during study (associative information). A conventional EEG analysis and a classifier analysis of this study were reported elsewhere (Borst et al., submitted). Currently, we are interested in finding out how many stages the subjects went through while determining a correct response. A new method is demonstrated for identifying processing stages in a task. Since the 1860s cognitive scientists have used different methods to identify processing stages, usually based on reaction time (RT) differences between conditions. To overcome the limitations of RT-based methods we used Hidden Markov Models (HMMs) to analyze EEG data. The HMMs indicate for how many stages there is evidence in the data, and how the durations of these stages vary with experimental condition. This method was applied to an associative recognition task in which associative strength and target/foil type were manipulated. The HMM-EEG method identified six different processing stages for targets and re- paired foils, whereas four similar stages were identified for new foils. The duration of the third, fifth and sixth stage varied with associative strength for targets and re-paired foils. We present an interpretation of the identified stages, and conclude that the method can provide valuable insight in human information processing. Keywords: EEG; HMM, processing stages. Introduction One of the main goals of cognitive science is to understand how humans perform tasks. To this end, scientists have long tried to identify different processing stages in human information processing. The first to do this in a systematic manner was probably Franciscus Donders. Almost 150 years ago, Donders proposed a method to measure the duration of cognitive stages (1868). By subtracting the RTs of two tasks that were hypothesized to share all but one processing stage, the duration of that stage could be calculated. A strong – and often problematic – assumption of Donders' subtractive method is the idea that it is possible to add an entire stage without changing the duration of other stages. To test whether different stages exist in the first place, Sternberg proposed the additive-factor method (1969). Although Sternberg overcame a limitation of Donders' method, the additive-factors method has its own drawbacks: it can only indicate the minimum number of stages in a task and it does not yield dur

展开