A computational model of short - term recognition and recall

In this study, a connectionist model for serial recall– Serial-Order in the Box (SOB)(Lewandowsky & Farrell, 2008; Oberauer, Lewandowsky, Farrell, Jarrold, & Greaves, 2012) – is extended for explaining data from short-term recognition. The main motivation behind the extension is to create a computational model which can simulate both short-term recall tasks and short-term recognition task, which in turns bridging both tasks under the same memory representation. In SOB, the memory representation consists of bindings between items and contexts (i.e. position markers of serial positions). The structure of the network and encoding process are kept unchanged for the recognition model, while the retrieval process for the recognition task differs from the serial-recall task. The retrieval process is modeled as comparing the probe to the memory content retrieved from the context, and the context used for retrieving the memory content is retrieved by activating the probe and deblurred through the recognition process. At the beginning of the retrieval process, the context used for retrieval is noisy because of the superposition between bindings, and it is subsequently sharpen to the context which is most strongly associated to the probe. Thus, the retrieved memory content is a mixture of all the memory items at the beginning and then is gradually narrowed down to the memory content which is most similar to the probe. The recognition model is able to simulate the set-size effect, the serial-position effect, and the speed-accuracy trade-off in both Sternberg’s memory scanning task. The model is also able to simulate the expectation effect for the incoming task, the local recognition task, and the performance from continuous stimulus recognition task. This is also the first computational model explaining both recognition and serial recall of information in working memory.