sign in sign up
<<
Home , Memory consolidation, Recall (memory)

TICS 1685 No. of Pages 4

Forum making the less shown to enhance memory retention [3,4]. Retrieval as a Fast dependent and more readily accessible Critically, we propose that the neural reac- in the future. We explore theoretical links tivation of recently acquired , as Route to Memory between retrieval and offline consolida- triggered online by incomplete reminders Consolidation tion, describe some key evidence in sup- (pattern completion), promotes long-term port of a shared mechanism, draw retention in a way similar to offline replay. James W. Antony,1,* parallels between this proposal and other 2 forms of rapid consolidation, and outline We argue that retrieval and can qual- Catarina S. Ferreira, predictions for future research (Box 1). itatively transform memories in at least two 1 Kenneth A. Norman, and distinct ways: byintegratingnew memories 2, Maria Wimber * Retrieval as a Rapid into preexisting neocortical knowledge Consolidation Event structures and by adaptively differentiating Retrieval-mediated is a Extensive evidence from and memories (i.e., reducing their neural over- powerful way to make memories patients shows that hippocampal lap) so[106_TD$IF] as to minimize competition between last, but its neurocognitive mecha- damage affects the formation of new overlapping memories. From a computa- nisms remain unclear. We propose declarative memories while leaving remote tional perspective, both the integration and that retrieval acts as a rapid con- memories (at least partially) intact [2].An differentiation effects can be explained by fl solidation event, supporting the in uential computational model [Comple- the tendency of retrieval to be imprecise; mentary Learning Systems (CLS)] [2] sug- that is, to coactivate memories that are creation of adaptive hippocampal- geststhatthehippocampusandneocortex semantically or episodically linked to the –neocortical representations via act synergistically to allow new learning target memory [5]. Repeated imprecise ‘ ’ the online reactivation of associa- while preserving old information. Specifi- reactivations in hippocampal–neocortical tive information. We describe par- cally, the learns slowly and spe- circuits afford an opportunity to integrate allels between online retrieval and cializes in storing the statistical structure of an initially hippocampus-dependent mem- offline consolidation and offer experiences. The hippocampus learns ory into the coactivated neocortical knowl- testable predictions for future quickly and specializes in rapidly edge structures, similar to replay events research. and binding together new cortical associ- during NREM sleep (Figure 1). According ations. Repeated interactions between the to [5], the of learning driven by coac- Introduction two systems allow new information to tivation depends on how strongly memo- For over a century, have slowly shape neocortical representations.[105_TD$IF] ries are activated: strong coactivation of known that repeatedly and actively retriev- If the hippocampus is damaged before memories leads to integration of those ing information from memory, as opposed enough hippocampal–neocortical interac- memories, whereas moderate activation to restudying the same information, tions can occur, long-term memory will be of competing memories triggers their strongly enhances long-term retention impaired. These ideas constitute systems- adaptive weakening [6] and pushes [1]. The benefits of retrieval-mediated level consolidation, or the process by retrieved and competing memories apart learning (also known as the ‘’) which newly acquired information is trans- in representational space [7],[107_TD$IF] leaving the hold across a wide variety of materials and formed into a stable, long-term memory retrieved memory in a distinct, accessible testing formats and remain evident across representation [3]. state forfuture (Figure 1).Importantly, much of the lifespan [1]. No mechanistic restudy (i.e., simple re-exposure to a com- framework exists to date integrating these The gradual transformation that a memory plete, previously stored memory) does not behavioral findings with the growing litera- undergoes during systems-level consoli- share these computational characteristics ture on the neural basis of learning and dation is promoted by the memory’s [6,7]. Restudy may re-impose the memo- memory [1]. Here we attempt such an repeated offline reactivation (‘replay’)in ry’s original pattern onto the hippocampus explanation. hippocampal–neocortical circuits. Reacti- and neocortex, causing some strengthen- vations during non-rapid eye movement ing of the original trace. However, because (NREM) sleep arguably play a unique role restudy triggers less coactivation of related In short, we propose that retrieval acts as in embedding information in the neocortex, memories it does not adaptively shape the a fast route to . facilitated by low cholinergic activity and hippocampal–neocortical memory land- Specifically, we propose that retrieval coordinated oscillatory interactions scape in the same way as active retrieval. integrates the memory with stored neo- between the hippocampus and neocortex cortical knowledge and differentiates it [4]. Replay occurring during both post- So far we have outlined the similarities from competing memories, thereby learning wakeful rest and sleep has been between retrieval and offline consolidation

Trends in Cognitive Sciences, Month Year, Vol. xx, No. yy 1 TICS 1685 No. of Pages 4

on a theoretical, computational level. We hippocampal representations to access memories from retroactive interference. next examine behavioral and neural par- the memory. Retrieval-mediated memory Including a period of sleep (as opposed allels between memory retrieval effects boosts should thus be most evident to wake) between two learning sessions (indexed by differences in retrieved versus whenever hippocampal traces are weak reduces retroactive interference [3], pre- restudied information) and consolidation and recall is relatively more dependent on sumably due to offline reactivation. effects (indexed by sleep versus wake the neocortex. Consistent with this Intriguingly, including retrieval practice intervals) that empirically support our notion, testing effects are strongest at between the[108_TD$IF] learning of two lists of words rapid consolidation view. long delays of several days to weeks similarly reduces retroactive interference [1], when recall supposedly relies more [8], consistent with a rapid online consoli- Similarities between Retrieval and heavily on neocortical traces than at short dation mechanism. Consolidation delays. If retrieval rapidly embeds a memory in the Sleep has been shown to globally neocortex, future retrievals of this mem- Online retrieval and offline reactivation enhance long-term retention but seems ory can utilize neocortical in addition to also share the property of protecting to preferentially benefit salient, rewarded,

Before Types During Aer retrieval/restudy of reacvaons retrieval/restudy retrieval/restudy

Retrieval Recent + (target) + Previous + +

(compeng) – + + + –

NC Restudy

+ HC + +

Suitcase/phone/taxi in NC

Suitcase/phone episodes in HC + / – Change in Informaon semancally related connecon weights to suitcase/phone in NC Acvaon strength Corcal connecons (weak/strong) during retrieval/restudy

Figure 1. Neural[103_TD$IF] Network Changes via Retrieval Compared with Restudy. Imagine you form a new memory of leaving your phone in a taxicab. Also imagine that 2 weeks before this episode you needed a taxi because you had a very heavy suitcase with you. The two episodes have distinct, nonoverlapping representations in the hippocampus (HC) but are connected via the shared element ‘taxi’ in the neocortex (NC). The NC also holds representations of memories that are loosely connected with phones and suitcases. Retrieval of the taxi–phone episode will activate the core representations in the HC and NC but will also cause coactivation of related information including the suitcase and knowledge related to phones. Strongly active memories and their connections (blue elements) will be strengthened (indicated by +), supporting memory integration. Moderately coactivated information will be weakened (indicated by À), supporting differentiation, while no change in connection strength occurs for weakly active memories. By comparison, restudy reinstantiates the cortical representations of only those elements that are explicitly presented; connections between these elements will be strengthened but the memory will not be integrated with or differentiated from other stored knowledge.

2 Trends in Cognitive Sciences, Month Year, Vol. xx, No. yy TICS 1685 No. of Pages 4

or emotional information containing future neocortical representations without via fast mapping or retrieval. Indeed, fast utility [3,4]. Assuming that prioritized infor- and sleep. Importantly, two other learning mapping and retrieval share common ele- mation is ‘tagged’ to receive greater off- paradigms suggest that rapid cortical ments of active inference and knowledge line processing during sleep, retrieval learning is possible. First, research on activation and may thus exert their benefi- could in theory interact with the consoli- schema-based learning demonstrates cial effects on long-term retention via the dation process in two different ways. If that information consistent with prior same mechanism. retrieval effectively tags information to knowledge (schemas) can become hip- receive greater subsequent offline proc- pocampus independent and embedded Concluding Remarks essing during sleep, retrieved memories in the neocortex surprisingly quickly and We propose that retrieval stabilizes mem- should benefit more from sleep than without intervening sleep [13]. As argued ories via mechanisms similar to those that restudied memories. By contrast, if the above, we assume that retrieval rapidly occur during sleep and offline consolida- benefits of retrieval occur online, as pro- consolidates memories in a similar way, tion periods. We acknowledge that sleep posed by our rapid consolidation frame- by coactivating related knowledge struc- has unique characteristics that are not work, sleep should offer greater relative tures to facilitate neocortical integration. shared with wake retrieval modes. At benefits to restudied information because present it is unclear to what degree reac- retrieved information has already been Second, in a related paradigm called ‘fast tivations during sleep are imprecise and consolidated online. In fact, restudied mapping’ individuals rely on prior knowl- associated with the same strengthening information does[109_TD$IF] benefit more from sleep edge to learn new information by active and weakening dynamics as reactivations than retrieved information [9], underscor- inference [14]. Intriguingly, patients with during wake. Nevertheless, the rapid con- ing the idea that sleep offers relatively hippocampal amnesia can learn new solidation view of retrieval makes testable more benefits to information that has objectsmoreeffectivelyusingfastmapping predictions for future research (Box 1) and not already been stabilized via online con- than a simple instruction to encode explic- will hopefully stimulate new research to solidation during retrieval. itly [14], suggesting that fast mapping bridge the existing gap between cognitive arises from a faster neocortical embedding and neuroscientific investigations of Systems consolidation accounts [2] state process [15]. Associations learned via fast memory modification. that initially hippocampus-dependent mapping (compared with explicit encod- fi [10_TD$IF] memories become gradually incorpo- ing) also bene t less from sleep [15], very Acknowledgments rated into the neocortex for stable long- similar to associations learned via active This work was supported by the CV Starr Fellowship term . If retrieval accelerates con- retrieval (compared with restudy) [9]. These (awardedto J.W.A.),NIMHR01MH069456(awardedto solidation, it should produce neural findings imply that offline enhancement is K.A.N.), and ESRC grant ES/M001644/1 (awarded to changes similar to those with offline con- reduced if cortical integration occurs online M.W.). solidation. In , connectivity between the hippocampus and the ven- tromedial (vmPFC), a key region in systems consolidation, increases after sleep [10] and retrieval practice but not after restudy [11]. Simi- larly, it has been shown in rodents that repeated reactivations support the crea- Box 1. Predictions and Open Questions for Future Research tion of memory traces that can be  fMRI studies should find a relative increase of vmPFC and decrease of hippocampal contributions when probing previously retrieved compared with restudied memories. accessed independently of the hippo-  Retrieval benefits (i.e., testing effects) should be associated with representational differentiation and campus [12], suggesting that online reac- competitor weakening [5,6][104_TD$IF] . tivation resembles offline consolidation at  Offline reactivation can occur spontaneously but can also be induced by presenting cues previously a neural level. paired with learning [in a technique known as targeted memory reactivation (TMR)] [4]. As cognitive control is absent during sleep, does TMR result in coactivation of related associations in a manner resembling retrieval during wake [5,6][104_TD$IF] ? Can Neocortical Representations  If repeated reactivations via retrieval help to embed memories in the neocortex and make them less Be Modified Rapidly? hippocampus dependent, this transformation should have observable effects on behavior. For example, While most frameworks propose that sys- retrieval, like sleep, should promote insight, inference, and generalization, more so than restudy [4], and may reduce context dependency during future recall. tems consolidation happens slowly via  Interactions between the hippocampus and neocortex should be most effective when capitalizing on hippocampal–neocortical interactions preexisting neocortical schemas [13]. If retrieval-mediated learning depends on such interactions, its over time [2], we propose that retrieval benefits will be reduced for novel materials that do not have preexisting neocortical representations (e.g., promotes the rapid development of nonsense drawings or syllables).

Trends in Cognitive Sciences, Month Year, Vol. xx, No. yy 3 TICS 1685 No. of Pages 4

1Princeton Neuroscience Institute, Princeton University, 3. Dudai, Y. et al. (2015) The consolidation and transforma- recall of retrieved items unaffected. J. Exp. Psychol. Learn. – – Princeton, NJ 08544, USA tion of memory. 88, 20 32 Mem. Cogn. 40, 1568 1581 2School of Psychology, University of Birmingham, 4. Antony, J.W. and Paller, K.A. (2017) Hippocampal contri- 10. Gais, S. et al. (2007) Sleep transforms the cerebral trace of Birmingham, B15 2TT, UK butions to declarative memory consolidation during sleep. declarative memories. Proc. Natl. Acad. Sci. U. S. A. 104, In The Hippocampus from Cells to Systems (Hannula, D.E. 18778–18783 and Duff, M.C., eds), pp. 245–280, Springer *Correspondence: 11. Wing, E.A. et al. (2013) Neural correlates of retrieval-based 5. Norman, K.A. et al. (2007) A neural network model of memory enhancement: an fMRI study of the testing effect. [email protected] (J.W. Antony) and retrieval-induced . Psychol. Rev. 114, 887–953 Neuropsychologia 51, 2360–23670 [email protected] (M. Wimber). 6. Wimber, M. et al. (2015) Retrieval induces adaptive for- 12. Lehmann, H. et al. (2009) Making context memories inde- http://dx.doi.org/10.1016/j.tics.2017.05.001 getting of competing memories via cortical pattern sup- pendent of the hippocampus. Learn. Mem. 16, 417–420 – pression. Nat. Neurosci. 18, 582 589 13. Tse, D. et al. (2007) Schemas and memory consolidation. References 7. Hulbert, J.C. and Norman, K.A. (2015) Neural differentia- Science 316, 76–82 tion tracks improved recall of competing memories follow- 14. Sharon, T. et al. (2011) Rapid neocortical acquisition of 1. Rowland, C.A. (2014) The effect of testing versus restudy ing interleaved study and retrieval practice. Cereb. Cortex long-term arbitrary associations independent of the hippo- – on retention: a meta-analytic review of the testing effect. 25, 3994 4008 campus. Proc. Natl. Acad. Sci. U. S. A. 108, 1146–1151 – Psychol. Bull. 140, 1432 1463 8. Potts, R. and Shanks, D.R. (2012) Can testing immunize 15. Himmer, L. et al. (2016) Sleep-mediated memory consoli- 2. McClelland, J.L. et al. (1995) Why there are complemen- memories against interference? J. Exp. Psychol. Learn. dation depends on the level of integration at encoding. – tary learning systems in the hippocampus and neocortex: Mem. Cogn. 38, 1780 1785 Neurobiol. Learn. Mem. 137, 101–106 insights from the successes and failures of connectionist 9. Bäuml, K.-H.T. et al. (2014) Sleep can reduce the testing models of learning and memory. Psychol. Rev. 102, effect: it enhances recall of restudied items but can leave 419–457

4 Trends in Cognitive Sciences, Month Year, Vol. xx, No. yy