تعمیم حافظه اپیزودیک در سراسر زمان: مسیر برای کسب دانش معنایی

The storage of input regularities, at all levels of processing complexity, is a fundamental property of the nervous system. At high levels of complexity, this may involve the extraction of associative regularities between higher order entities such as objects, concepts and environments across events that are separated in space and time. We propose that such a mechanism provides an important route towards the formation of higher order semantic knowledge. The present study assessed whether subjects were able to extract complex regularities from multiple associative memories and whether they could generalize this regularity knowledge to new items. We used a memory task in which subjects were required to learn face-location associations, but in which certain facial features were predictive of locations. We assessed generalization, as well as memory for arbitrary stimulus components, over a 4-h post-encoding consolidation period containing wakefulness or sleep. We also assessed the stability of regularity knowledge across a period of several weeks thereafter. We found that subjects were able to detect the regularity structure and use it in a generalization task. Interestingly, the performance on this task increased across the 4hr post-learning period. However, no differential effects of cerebral sleep and wake states during this interval were observed. Furthermore, it was found that regularity extraction hampered the storage of arbitrary facial features, resulting in an impoverished memory trace. Finally, across a period of several weeks, memory for the regularity structure appeared very robust whereas memory for arbitrary associations showed steep forgetting. The current findings improve our understanding of how regularities across memories impact memory (trans)formation.

Episodic memory refers to memory for events and situations, organized in an autobiographical stream, and rich in contextual information (Tulving, 1983). The hippocampus is thought to be crucial for the encoding and retrieval of such vivid memories (Burgess et al., 2002, McClelland et al., 1995, Nadel et al., 2000, Squire, 1992 and Tulving and Markowitsch, 1998). Semantic memory, on the other hand, refers to general knowledge about the world (McClelland and Rogers, 2003, Moscovitch et al., 2005 and Patterson et al., 2007) and is stored in networks comprising widespread cortical regions (Binder et al., 2009 and Martin and Chao, 2001). Though there may be various routes towards creating semantic memories, the acquisition of complex semantic knowledge frames appears to involve episodic memory (Rosenbaum et al., 2001 and Moscovitch et al., 2005) and an intact hippocampus (Bayley and Squire, 2005, Hayman et al., 1993 and Manns et al., 2003; but see Gardiner, Brandt, Baddeley, Vargha-Khadem, & Mishkin, 2008 for a critical viewpoint). It has, in fact, been proposed that semantic memory is extracted from episodic memories through a hippocampo-cortical dialogue (Battaglia and Pennartz, 2011, McClelland et al., 1995 and Meeter and Murre, 2005). According to the pertaining accounts, event memory would initially require a highly plastic memory system, the hippocampus, to quickly store the conjunction of different event components represented in widespread cortical circuits. Reinstatement of overlapping event memories, facilitated by the hippocampus, would allow less plastic extrahippocampal systems, to gradually discover the structure across ensembles of experiences. Thus, memories for statistical regularities, which are important for predicting behaviour in the long run, would acquire a neural representation that is relatively resistant to decay. On the other hand, arbitrary associations would tend to be forgotten at relatively high rate, consequent to fast overwriting of neural patterns in the hippocampus. This general framework accounts for a large spectrum of neurophysiological and behavioural observations regarding mammalian learning. In particular, the time-limited involvement of the hippocampus in (many forms of) memory (o.a. Kim & Fanselow, 1992) and the occurrence of temporally graded amnesia following brain damage (Brown, 2002, Ribot, 1881 and Squire et al., 1975), especially when such damage involves the hippocampus (Manns et al., 2003, Murre et al., 2006, Scoville and Milner, 1957 and Squire et al., 2001). More recent support comes from studies on system-level consolidation, showing that hippocampus-dependent memories, over time, acquire a more cortically based representation (Frankland et al., 2004, Paz et al., 2007, Takashima et al., 2006 and Takashima et al., 2009). Another study directly compared hippocampally and extrahippocampally stored aspects of events, and showed that memories preferentially lose hippocampus-dependent configurational components as they age (Talamini & Gorree, 2012). Hippocampal lesion studies in rodents suggest that such contextual impoverishment coincides with hippocampal disengagement from retrieval (Wiltgen et al., 2010 and Winocur et al., 2009). Several aspects of the model have thus been supported by experimental evidence, including hippocampo-neocortical memory recoding over time, and a parallel loss of memories' configural complexity, leading to memory semantization. On the other hand, the proposed hippocampus-mediated regularity extraction across episodic memories has long remained hypothetical.1 A recent study from our lab generated initial support for this notion (Sweegers, Takashima, Fernández, & Talamini, 2013). In this study, subjects were required to learn associations between faces and screen locations. Half of the associations harboured complex regularities, in that facial features were predictive of screen locations; the other half did not. Importantly, regularities could only be extracted over hippocampus-encoded, associative aspects of the items. We will henceforward use the term ‘associative regularity extraction’ to denote this type of regularity extraction. The results indicated enhanced memory encoding and retention when associative regularities could be extracted. Moreover, the build-up of general knowledge regarding regular associations involved the coordinated activity of the hippocampus and mediofrontal regions, and the build-up of a functionally interconnected (neo)cortical network. These findings suggest that cross episodic regularities gain preferential access to system-level consolidation processes, over arbitrary episodic memory components, leading to preferential maintenance of such regularities and, therewith, the build-up of general knowledge. Significant questions, however, remain. For one, our previous study (Sweegers et al., 2013) did not test whether acquired regularity knowledge could be generalized to new situations (c.q. new exemplars). Thus, definitive proof that flexibly applicable, general associative knowledge can be extracted across episodes is still lacking. Ulterior questions regard the temporal dynamics of associative regularity extraction. In particular, does such extraction occur largely at the moment when a new episode occurs that holds overlap with a stored one, or might it proceed after encoding? There is some indication that the extraction of regularities across temporally spaced events may further develop during offline consolidation processes. In the pertaining studies sets of items were encoded, that – unbeknownst to the subjects – carried a hidden structure or grammar (Durrant et al., 2012, Ellenbogen et al., 2007, Gómez et al., 2006 and Tamminen et al., 2012). While subjects immediately grasped some parts of such underlying structures, the extraction of additional relations evolved across time. In these studies, however, the encoding of the individual items may not have been crucially dependent upon hippocampal functioning. If generalization of associative regularities does continue post-encoding, this might occur differentially during the wake and sleep state. Indeed, sleep has been shown to contribute importantly to both the consolidation (Gais et al., 2006, Payne et al., 2008, Plihal and Born, 1997 and Talamini et al., 2008) and the reorganization of memories (Djonlagic et al., 2009, Ellenbogen et al., 2007, Lau et al., 2011 and Orban et al., 2006; see Lewis & Durrant, 2011 for a model on the stabilization of regularities during sleep). Thus, it is possible that the extraction of associative regularity knowledge might also occur preferentially during such an ‘offline’ processing period. A final question is what happens to arbitrary aspects of episodes harbouring regular associations (i.e. episodic components that do not reflect any regularity). Are these lost at the average hippocampal decay rate, maintained preferentially, along with the regular association(s) in the episode or, rather, lost preferentially, consequent to some trade-off between the storage of regularities and of unique episodic details? The present study was designed to investigate the extraction of regularities across memories that require the hippocampus to be stored. To answer the particular questions that were raised in the paragraphs above, we used a face-location memory task similar to the one used in our previous study (Sweegers et al., 2013). In half of the associations, combinations of facial features were predictive of the face's location (e.g. slender faces with wrinkles and headwear appeared in one particular location); the other half of the faces were randomly assigned to a location. Importantly, the regularity structure in the material was of such complexity that it could only be extracted slowly and gradually, across many exemplars. Shortly following acquisition, and after a 4-h delay containing either a nap or quiet waking, face-cued location retrieval was tested. At these same times, we also assessed generalization of extracted regularity knowledge to new face-location exemplars and memory for arbitrary (regularity-irrelevant) facial features. Finally, longer-term memory for arbitrary and stable associations was tested at one month post-encoding.

The present study shows that when memories contain associative regularities, encoding and consolidation mechanisms appear particularly devoted to storing these regularities, at the cost of arbitrary memory components. We propose that the stabilization of associative regularities across memories may be one route toward the build-up of our large semantic knowledge base. Such a mechanism seems highly adaptive, given limited memory capacity and the long-term relevance of associative regularities for predicting the more complex aspects of our environment and guiding behaviour accordingly.