Emergence of Semantic Memory through Sequential Event Prediction and Its Role in Episodic Future Thinking: A Computational Exploration

Emergence of Semantic Memory through Sequential Event Prediction and Its Role in Episodic Future Thinking: A Computational Exploration Yuichi Ito (ito.yuichi@nagoya-u.jp) Taiji Ueno (taijiueno7@gmail.com) Shinji Kitagami (kitagami@cc.nagoya-u.ac.jp) Jun Kawaguchi (kawaguchijun@nagoya-u.jp) Department of Psychology, Graduate School of Environmental Studies, Nagoya University, Furo-cho, Chikusa-ku, Nagoya City, Aichi 4648601, JAPAN precedes retrieval of time-specific episodic information. The role of semantic memory is also supported by neuropsychological data from patients with semantic dementia (SD) (Irish et al., 2012). These patients are impaired on tasks that probe conceptual knowledge of things (words, object, etc.), but their episodic memory (especially, about recent events) are relatively preserved (Irish et al., 2011). Irish et al. (2012) revealed that their episodic future thoughts lacked details relevant to the events cued by investigators. For example, when asked to talk about a future dinner, an SD patient might suddenly change topic and talk about his wife or past events. In other words, SD patients’ future simulations tend to transgress the boundary of the contexts cued by investigators. However, it has yet to be clarified why and how these different types of representations contribute to the simulation of future events. An implemented computational model is a useful approach on this issue (e.g., Botvinick & Plaut, 2004; Elman, 1990). Any computational modelling of a complex higher-order cognitive function requires a set of working assumptions and simplifications. It is noteworthy that simulation of future events involves the self-generation of successive internal event predictions. Taking a simulation of tomorrow morning as an example, one might first envision waking up in your bedroom, followed by an image of a next plausible event such as leaving the room, and finally one might imagine washing his/her face. In other words, our working assumption is that mental simulation requires a mechanism that allows sequential prediction of an event after the previously self-generated event in a plausible order. Interestingly, a seminal work of Rumelhart et al. (1986) mentioned this idea more than a quarter-century ago: Abstract This study aimed to clarify the mechanism underlying episodic future thinking, which refers to the ability to generate prospective events in a specific time/location/context. Given that episodic future thinking involves generating predictions in a plausible order from previous internal predictions, we hypothesized that knowledge of sequential event prediction should underlie episodic future thinking. A parallel-distributed processing model was trained to predict the next event in the training sequence. After training, the model used the acquired knowledge to repeatedly self- generate event sequences (i.e., the model predicts the next event, and this prediction then forms the input of the next trial which in turn will trigger the next prediction). The resultant event sequences captured the episodic future thinking of normal participants and that of neurological patients when the model was lesioned. Moreover, the nature of knowledge acquired after training for sequential prediction of external events reflected that of episodic memory, schema-like knowledge and semantic memory, all of which have been found to contribute to episodic future thinking by past studies. Keywords: episodic future thinking; semantic memory; parallel distributed processing model; sequential prediction Introduction We can mentally simulate future events that are likely to happen in a specific time and place (e.g., “We’ll go to that Indian restaurant for lunch. Upon arrival, a young waiter will say hello and show us to our table”). This cognitive function is termed episodic future thinking (e.g., Schacter, Addis & Buckner, 2008), and past studies have investigated the role of various types of knowledge within this mental simulation process. These include episodic memory (Hassabis et al., 2007), autobiographical memory (D’Argembeau & Mathy, 2011), schema/schemata representations (general knowledge database about a location/context where a mental simulation is projected) (Berntsen & Bohn, 2010), semantic memory 1 (Irish et al., 2012), and so on. For example, D’Argembeau and Mathy (2011) argued that construction of future event representations typically involves gradual conversion from general to more specific information such that access to general knowledge (autobiographic memory, and schema) Now, suppose that the world events did not happen. It would be possible to take the output of the mental model and replace the stimulus inputs from the world with inputs from our model of the world. In this case, we could expect that we could run a mental simulation and imagine the events that would take place in the world when we performed a particular action. Sequential predictions and various knowledge Once we formulate a mental simulation of future events in terms of sequential predictions (NB. We do not mean episodic future thinking is equal to sequential predictions), then we can explain why the experimental studies above found correlations between episodic future thinking and various types of knowledge (episodic memory, In literature, general knowledge of an event (i.e., schema/script) and semantic memory of an event are sometimes used as synonymously and/or the latter is the source of the former (Berntsen & Bohn, 2010; Schacter et al., 2008).