Knowledge And The Conscious State Of Knowledge In Implicit Learning

Consciousness is taken as one of the most mysterious problems in for human beings in 21st century. Among all the areas referring to consciousness, implicit learning was a popular one. There are Artificial Grammar Language (AGL) Paradigm, Serial Prediction (SP) Paradigm, and others that worked to explore unconscious knowledge. Instead of using knowledge itself as dependent variables, SP related changes of reaction times to knowledge, which made it even harder to conclude.Here in this study, we are interested in and focuse on the measurement of consciousness and unconsciousness, by combining protocol report, guess criterion, zero-correlation criterion and structural knowledge.The study mainly explored knowledge and its related conscious state. It was composed of two parts:the first one adopted AGL and finished three experiments; the second adopted SP and finished two experiments. These experiments tried to answer the questions about what has been learnt under both conditions, and what kind of conscious state the knowledge were in. Results showed:(1) By comparing their knowledge of chunks (high chunk strength vs. low chunk strength) with that of grammars (ruled vs. non-ruled) under both recognition and classification condition, we found that recognition was mainly influenced by chunks, which was stable through less and more learning contexts; instead, classification was related to both chunks and grammars, which could be improved through training.(2) We adopted classical Serial Prediction task by asking participants to predict either self-initiated or procedure-initiated, which means they could choose to predict or not in the former situation, while they could not in the latter. There were some interesting results:in self-initiated situation, participants could predict items that were decided by a fixed sequence (sequential knowledge), while they were not sensitive at all to the probability rules; On the other hand, in procedure-initiated situation, participants did quite well on both sequential and probability knowledge.(3) According to zero-correlation criterion, we found that in experiment 2 and experiment 3, knowledge about regularities were positively related to subjects' confidence rating, which meant that knowledge was in some kind of conscious state. However, ROC showed no reasonable change of either hit or false alarm, which contradicted with the assumption Snodgrass (2002) that the distribution of conscious knowledge should be Guassian, in which both hit and false alarm rate were running the opposite way from confidence.(4) We adopted confidence rating by asking people to bet on their knowledge instead of confidence rating. Results showed that the correlation between the betting and their performance were largely affected by individual difference, for instance, the tendency to take risk. The same situation with experiment 3 happened to experiment 5 when average performance were taken into account, that correlation between confidence and prediction were positive while ROC showed negative answer to consciousness. However, zero-correlation was satisfied when we exclude risky factors. Case analysis on one participant who had experience on gamble and made the most wages on his knowledge showed that his prediction was well connected to probability (the "rules"). We could make an inference that the unconscious knowledge that was confounded with conscious by conservative subjects was cleared up for the one who was playing a risky game.(5) By forcing participants to make structural attribution choice (Dienes et al., 2008), i.e. Are there judgment based on randomly guess, intuition, familiarity or abstract rules? Results turned out to be consistent with prior experiments by Scott & Dienes (2009), that knowledge about grammars was generally based on the feeling of familiarity, although ROC again showed difference between familiarity in our experiment and that in his. More than that, knowledge about chunks was mostly based on their randomly guess, as well as intuition. The differences were worthy of considering.(6) In order to find out the internal process of implicit learning, we used Artificial Neural Networks to simulate the classification behavior of human beings. Results showed that network as simple as BP (Backward Propagation) Feedforward modal was enough to make good simulation, suggesting that both the memory buffer (in Kuhn & Dienes,2008) and recurrent context (in Cleeremans et al.,1998) were unnecessary in our situation. BP behaved even better than human beings.