We agree with Gilead et al. on the necessity of a hierarchical organization of abstract representations in long-term memory (LTM) for the efficient construction of new mental representations. Although this aspect is merely mentioned in the target article, it is closely related to a core assumption of predictive brain theories. Specifically, as a predictive machine, the brain should encode new information by “explaining away” discrepancies between incoming signals and top-down predictions based on prior knowledge (internal models) acquired in the past (Clark Reference Clark2013). Then, discrepancies can be used as bottom-up error signals to update internal models. It follows that a deeper integration between Gilead and colleagues’ conceptualization and the predictive brain framework requires to specify abstractions in terms of updated representations of previously stored abstractions. Moreover, new abstractions need then to be consolidated and integrated into the complex and hierarchically structured representations stored in LTM to be used for future acts of abstraction.
Based on these theoretical considerations, we want to highlight the importance of making a distinction between online and offline abstraction processes. Online abstractions can be seen as short-term inferential processes that, similarly to the ones mainly described in the target article, lead to new (updated) mental representations by integrating evidence – whose content is driven by ongoing internal/external experience – with prior abstractions retrieved from LTM. The short-term status of online abstractions is fundamental to avoid the simple “hoarding” of representations in LTM, which would lead to a not efficient organization of internal models. Conversely, efficient integration of new representations with pre-existing abstracta requires a deep reorganization of pre-existing representations to maintain coherence in such internal models and to ensure their predictive power by preventing redundancy and overfitting (Pezzulo et al. Reference Pezzulo, Kemere and van der Meer2017). Therefore, some key questions not discussed by Gilead and colleagues need to be addressed: How and when are abstract representations consolidated into long-term memory? And how are they integrated with pre-existing abstracta? Here, we posit that the complexity of such reorganization requires offline acts of abstraction during which the brain – decoupled from ongoing experience and in the absence of new evidence – recursively generates and evaluates inferential updates to improve the hierarchical organization of the internal models (Pezzulo et al. Reference Pezzulo, Kemere and van der Meer2017). Grounding on compelling evidence that has emerged in the past two decades about the active role of sleep in (i) consolidating spatial, episodic, and semantic memories (Rasch & Born Reference Rasch and Born2013), (ii) facilitating the extraction of statistical regularities, gist, and overarching rules (Stickgold & Walker Reference Stickgold and Walker2013), and (iii) supporting the formation of internal predictive models (Lutz et al. Reference Lutz, Wolf, Hübner, Born and Rauss2018), sleep appears to be the ideal status where offline abstraction processes may take place.
More intriguingly, we propose that the sequential interplay between non-rapid eye movement (NREM) and REM sleep, which iteratively interleaves in 90-min cycles during nocturnal sleep, may be critical for the development of the representational hierarchy proposed by Gilead and colleagues by promoting both the consolidation and the updating (reorganization) of these internal models (Ballesio & Cellini Reference Ballesio and Cellini2019). Indeed, NREM and REM sleep show very different patterns of brain activity, which likely support complementary functions in structuring, consolidating, updating, and minimizing the complexity of internal models (Hobson & Friston Reference Hobson and Friston2012). In particular, NREM sleep may promote an initial consolidation of newly encoded information into pre-existing memories that are coherent with, and supportive of, our prior beliefs and knowledge (Oudiette & Paller Reference Oudiette and Paller2013). Compelling evidence indicates that in NREM2 and NREM3 the hippocampus controls neural replay in the cortex ensuring that memories related to a prior representation are replayed concurrently (Klinzing et al. Reference Klinzing, Niethard and Born2019). This overlapping replay seems to lead to the abstraction of shared properties although redundant information is pruned (Feld & Born Reference Feld and Born2017; Lewis et al. Reference Lewis, Knoblich and Poe2018; Stickgold & Walker Reference Stickgold and Walker2013). Once coded in the neocortex during NREM sleep, such representations might serve then, as concreta for further abstraction processes during REM sleep. Indeed, during REM a reduced hippocampal input to neocortex is coupled with a massive neural activity (driven by ponto-geniculo-occipital waves) that can randomly activate cortical representations. This highly active state might allow scanning memories to find meaningful regularities between events, and to generate and strengthen new abstractions between both cortical representations activated during the NREM and other, apparently not related, representations (Lewis et al. Reference Lewis, Knoblich and Poe2018; Llewellyn Reference Llewellyn2016). Moreover, it has been proposed that during REM the sleeping brain tends to simulate and rehearse fictive scenarios to produce prediction errors and, therefore, refine (i.e., update) the internal models before the next wakefulness (Friston et al. Reference Friston, Lin, Frith, Pezzulo, Hobson and Ondobaka2017b; Hobson et al. Reference Hobson, Hong and Friston2014).
Finally, we believe the different nature of neural activity patterns and associated cognitive processes observed during NREM and REM sleep may support the qualitative distinction between regular abstracta and predicators. According to Gilead and colleagues, regular abstracta can be seen as saturated representations that arise from regularity and/or shared properties between concreta. Conversely, predicators are unsaturated entities that behave like functions to create new associations between representations. This distinction emerges also in the abstraction processes described above. Indeed, during NREM abstractions emerge from the overlap of shared properties between representations, whereas in REM new associations between representations are instantiated. Therefore, beyond the qualitative difference between abstract representations, we propose that future studies may take advantage of investigating qualitative differences in abstraction processes such as those promoted by the different physiology and neural mechanisms that the brain experiences in NREM and REM sleep.
In sum, we believe that differentiating between online and offline (particularly sleeping) abstraction processes is fundamental for understanding how the representational hierarchy proposed in the target article is built up and optimized for prediction. Moreover, based on the different neural activities observed in NREM and REM sleep, we suggest that the identification of qualitative different abstraction processes (beyond the qualitative difference in representations) will offer a key to unveil the representational bases of cognition.
You have
Access
We agree with Gilead et al. on the necessity of a hierarchical organization of abstract representations in long-term memory (LTM) for the efficient construction of new mental representations. Although this aspect is merely mentioned in the target article, it is closely related to a core assumption of predictive brain theories. Specifically, as a predictive machine, the brain should encode new information by “explaining away” discrepancies between incoming signals and top-down predictions based on prior knowledge (internal models) acquired in the past (Clark Reference Clark2013). Then, discrepancies can be used as bottom-up error signals to update internal models. It follows that a deeper integration between Gilead and colleagues’ conceptualization and the predictive brain framework requires to specify abstractions in terms of updated representations of previously stored abstractions. Moreover, new abstractions need then to be consolidated and integrated into the complex and hierarchically structured representations stored in LTM to be used for future acts of abstraction.
Based on these theoretical considerations, we want to highlight the importance of making a distinction between online and offline abstraction processes. Online abstractions can be seen as short-term inferential processes that, similarly to the ones mainly described in the target article, lead to new (updated) mental representations by integrating evidence – whose content is driven by ongoing internal/external experience – with prior abstractions retrieved from LTM. The short-term status of online abstractions is fundamental to avoid the simple “hoarding” of representations in LTM, which would lead to a not efficient organization of internal models. Conversely, efficient integration of new representations with pre-existing abstracta requires a deep reorganization of pre-existing representations to maintain coherence in such internal models and to ensure their predictive power by preventing redundancy and overfitting (Pezzulo et al. Reference Pezzulo, Kemere and van der Meer2017). Therefore, some key questions not discussed by Gilead and colleagues need to be addressed: How and when are abstract representations consolidated into long-term memory? And how are they integrated with pre-existing abstracta? Here, we posit that the complexity of such reorganization requires offline acts of abstraction during which the brain – decoupled from ongoing experience and in the absence of new evidence – recursively generates and evaluates inferential updates to improve the hierarchical organization of the internal models (Pezzulo et al. Reference Pezzulo, Kemere and van der Meer2017). Grounding on compelling evidence that has emerged in the past two decades about the active role of sleep in (i) consolidating spatial, episodic, and semantic memories (Rasch & Born Reference Rasch and Born2013), (ii) facilitating the extraction of statistical regularities, gist, and overarching rules (Stickgold & Walker Reference Stickgold and Walker2013), and (iii) supporting the formation of internal predictive models (Lutz et al. Reference Lutz, Wolf, Hübner, Born and Rauss2018), sleep appears to be the ideal status where offline abstraction processes may take place.
More intriguingly, we propose that the sequential interplay between non-rapid eye movement (NREM) and REM sleep, which iteratively interleaves in 90-min cycles during nocturnal sleep, may be critical for the development of the representational hierarchy proposed by Gilead and colleagues by promoting both the consolidation and the updating (reorganization) of these internal models (Ballesio & Cellini Reference Ballesio and Cellini2019). Indeed, NREM and REM sleep show very different patterns of brain activity, which likely support complementary functions in structuring, consolidating, updating, and minimizing the complexity of internal models (Hobson & Friston Reference Hobson and Friston2012). In particular, NREM sleep may promote an initial consolidation of newly encoded information into pre-existing memories that are coherent with, and supportive of, our prior beliefs and knowledge (Oudiette & Paller Reference Oudiette and Paller2013). Compelling evidence indicates that in NREM2 and NREM3 the hippocampus controls neural replay in the cortex ensuring that memories related to a prior representation are replayed concurrently (Klinzing et al. Reference Klinzing, Niethard and Born2019). This overlapping replay seems to lead to the abstraction of shared properties although redundant information is pruned (Feld & Born Reference Feld and Born2017; Lewis et al. Reference Lewis, Knoblich and Poe2018; Stickgold & Walker Reference Stickgold and Walker2013). Once coded in the neocortex during NREM sleep, such representations might serve then, as concreta for further abstraction processes during REM sleep. Indeed, during REM a reduced hippocampal input to neocortex is coupled with a massive neural activity (driven by ponto-geniculo-occipital waves) that can randomly activate cortical representations. This highly active state might allow scanning memories to find meaningful regularities between events, and to generate and strengthen new abstractions between both cortical representations activated during the NREM and other, apparently not related, representations (Lewis et al. Reference Lewis, Knoblich and Poe2018; Llewellyn Reference Llewellyn2016). Moreover, it has been proposed that during REM the sleeping brain tends to simulate and rehearse fictive scenarios to produce prediction errors and, therefore, refine (i.e., update) the internal models before the next wakefulness (Friston et al. Reference Friston, Lin, Frith, Pezzulo, Hobson and Ondobaka2017b; Hobson et al. Reference Hobson, Hong and Friston2014).
Finally, we believe the different nature of neural activity patterns and associated cognitive processes observed during NREM and REM sleep may support the qualitative distinction between regular abstracta and predicators. According to Gilead and colleagues, regular abstracta can be seen as saturated representations that arise from regularity and/or shared properties between concreta. Conversely, predicators are unsaturated entities that behave like functions to create new associations between representations. This distinction emerges also in the abstraction processes described above. Indeed, during NREM abstractions emerge from the overlap of shared properties between representations, whereas in REM new associations between representations are instantiated. Therefore, beyond the qualitative difference between abstract representations, we propose that future studies may take advantage of investigating qualitative differences in abstraction processes such as those promoted by the different physiology and neural mechanisms that the brain experiences in NREM and REM sleep.
In sum, we believe that differentiating between online and offline (particularly sleeping) abstraction processes is fundamental for understanding how the representational hierarchy proposed in the target article is built up and optimized for prediction. Moreover, based on the different neural activities observed in NREM and REM sleep, we suggest that the identification of qualitative different abstraction processes (beyond the qualitative difference in representations) will offer a key to unveil the representational bases of cognition.
Acknowledgments
This work was supported by the University of Padova under the STARS Grants program to N.C. The present work was carried out within the scope of the research program “Dipartimenti di Eccellenza,” which is supported by a grant from MIUR to the Department of General Psychology, University of Padua.