Gilead et al. argue that the relationship between mental travel and abstraction is such that mental travel critically relies on abstraction, and that the function of abstraction is to support mental travel. I argue that in addition, mental travel – in particular, mind wandering – might facilitate abstraction, suggesting that the relationship between mental travel and abstraction as described by Gilead et al. might in fact be mutualistic. Abstraction is essential for making predictions, and critically relies on detecting invariance among experiences. For instance, based on my experience with rivers, I know that my feet get wet if I would step in one. This invariance is introduced by generalization across multiple instantiations of episodic experience, and is learnt from similarity and dissimilarity across them (Sloutsky Reference Sloutsky2003). I argue that mind wandering, which involves the spontaneous retrieval of episodic experiences, might help identify these similarities and dissimilarities, much like data augmentation in machine learning.
Mind wandering is a multidimensional construct that includes (but is not limited to) spontaneous mental travel (Christoff et al. Reference Christoff, Irving, Fox, Spreng and Andrews-Hanna2016). Despite its ubiquity – rates of up to 50% of the time have been reported (Kane et al. Reference Kane, Brown, McVay, Silvia, Myin-Germeys and Kwapil2007; Killingsworth & Gilbert Reference Killingsworth and Gilbert2010) – little is known about the function of this seemingly costly process. In recent work, Mills, Herrera-Bennett, Christoff and I proposed that the function of mind wandering might be to support episodic efficiency and semantic abstraction (Mills et al. Reference Mills, Herrera-Bennett, Faber, Christoff, Christoff and Fox2018). Specifically, we proposed the default variability hypothesis: mind wandering provides variability in mental content that helps to optimize the distinctiveness of episodic instantiations, which supports the extraction of invariant features of representations that ultimately lead to abstraction.
In brief, mind wandering can be characterized by its varying, dynamic content: the mind figuratively “wanders” from one thought to the next (Mills et al. Reference Mills, Raffaelli, Irving, Stan and Christoff2017; Reference Mills, Herrera-Bennett, Faber, Christoff, Christoff and Fox2018). These thoughts are often largely disjointed, although they might share one or more overlapping features (Faber & D'Mello Reference Faber and D'Mello2018). Take, for instance, the following example: When reading a text about chemical properties of water, a person might think about a beach near their house, followed by a thought about their job as a beach tagger during high school, followed by a thought about a person they used to like (from Faber & D'Mello Reference Faber and D'Mello2018). During this process, which we experience as a “train of thought,” one thought likely serves as a partial cue for the next (Faber & D'Mello Reference Faber and D'Mello2018). The process of retrieving a full memory from a partial or degraded cue is known as pattern completion, and is thought to be one of the key features of the human hippocampus (Marr Reference Marr1971). Indeed, recent work has shown that the hippocampus plays a critical role in spontaneous mental time travel (McCormick et al. Reference McCormick, Rosenthal, Miller and Maguire2018), as its role in spontaneous retrieval of memories from partial cues facilitates mental activity that transcends the here-and-now (Faber & Mills Reference Faber and Mills2018).
Importantly, the variability of content during mind wandering might support abstraction: by spontaneously retrieving a memory in a new context – either in reference to the external world or internal world – similarities across instantiations help identify the regularities necessary for abstraction. This process bears similarities with data augmentation in machine learning: diversity in data is increased without collecting new data by slightly modifying existing data, which are reused to train a model. Images, for instance, can be flipped, cropped, or partly occluded, which effectively adds noise that is useful for learning regularities across instances. A system that is learning to identify for example cats can benefit from being exposed to images that are flipped (cats can be viewed from different angles), cropped (the environment has little predictive value), or occluded (a particular feature of a specific cat might not generalize to all cats) to end up with a stable representation of “cats.”
In analogy, the (re)activation of (novel combinations of) episodic experiences in the context of an unrelated or tangentially related thought or physical environment adds noise that might be useful for identifying regularities across instantiations. The data augmentation induced by mind wandering might involve noise consisting of partial activations of an experience (similar to occlusion), and/or disrupted spatiotemporal contiguity induced by the novel internal or external environment (similar to cropping). This theory suggests a potential role for mind wandering – and mental travel more generally – in facilitating abstraction through data augmentation. Adding to Gilead et al.'s theory, the ideas laid out here imply that the relationship between mental travel and abstraction is in fact mutualistic: abstraction facilitates mental travel, and mental travel facilitates abstraction. Taking into consideration the potentially mutualistic nature of this relationship is critical to understanding both mental travel and abstraction, as well as to understanding the function of the seemingly costly cognitive process of mind wandering.
Gilead et al. argue that the relationship between mental travel and abstraction is such that mental travel critically relies on abstraction, and that the function of abstraction is to support mental travel. I argue that in addition, mental travel – in particular, mind wandering – might facilitate abstraction, suggesting that the relationship between mental travel and abstraction as described by Gilead et al. might in fact be mutualistic. Abstraction is essential for making predictions, and critically relies on detecting invariance among experiences. For instance, based on my experience with rivers, I know that my feet get wet if I would step in one. This invariance is introduced by generalization across multiple instantiations of episodic experience, and is learnt from similarity and dissimilarity across them (Sloutsky Reference Sloutsky2003). I argue that mind wandering, which involves the spontaneous retrieval of episodic experiences, might help identify these similarities and dissimilarities, much like data augmentation in machine learning.
Mind wandering is a multidimensional construct that includes (but is not limited to) spontaneous mental travel (Christoff et al. Reference Christoff, Irving, Fox, Spreng and Andrews-Hanna2016). Despite its ubiquity – rates of up to 50% of the time have been reported (Kane et al. Reference Kane, Brown, McVay, Silvia, Myin-Germeys and Kwapil2007; Killingsworth & Gilbert Reference Killingsworth and Gilbert2010) – little is known about the function of this seemingly costly process. In recent work, Mills, Herrera-Bennett, Christoff and I proposed that the function of mind wandering might be to support episodic efficiency and semantic abstraction (Mills et al. Reference Mills, Herrera-Bennett, Faber, Christoff, Christoff and Fox2018). Specifically, we proposed the default variability hypothesis: mind wandering provides variability in mental content that helps to optimize the distinctiveness of episodic instantiations, which supports the extraction of invariant features of representations that ultimately lead to abstraction.
In brief, mind wandering can be characterized by its varying, dynamic content: the mind figuratively “wanders” from one thought to the next (Mills et al. Reference Mills, Raffaelli, Irving, Stan and Christoff2017; Reference Mills, Herrera-Bennett, Faber, Christoff, Christoff and Fox2018). These thoughts are often largely disjointed, although they might share one or more overlapping features (Faber & D'Mello Reference Faber and D'Mello2018). Take, for instance, the following example: When reading a text about chemical properties of water, a person might think about a beach near their house, followed by a thought about their job as a beach tagger during high school, followed by a thought about a person they used to like (from Faber & D'Mello Reference Faber and D'Mello2018). During this process, which we experience as a “train of thought,” one thought likely serves as a partial cue for the next (Faber & D'Mello Reference Faber and D'Mello2018). The process of retrieving a full memory from a partial or degraded cue is known as pattern completion, and is thought to be one of the key features of the human hippocampus (Marr Reference Marr1971). Indeed, recent work has shown that the hippocampus plays a critical role in spontaneous mental time travel (McCormick et al. Reference McCormick, Rosenthal, Miller and Maguire2018), as its role in spontaneous retrieval of memories from partial cues facilitates mental activity that transcends the here-and-now (Faber & Mills Reference Faber and Mills2018).
Importantly, the variability of content during mind wandering might support abstraction: by spontaneously retrieving a memory in a new context – either in reference to the external world or internal world – similarities across instantiations help identify the regularities necessary for abstraction. This process bears similarities with data augmentation in machine learning: diversity in data is increased without collecting new data by slightly modifying existing data, which are reused to train a model. Images, for instance, can be flipped, cropped, or partly occluded, which effectively adds noise that is useful for learning regularities across instances. A system that is learning to identify for example cats can benefit from being exposed to images that are flipped (cats can be viewed from different angles), cropped (the environment has little predictive value), or occluded (a particular feature of a specific cat might not generalize to all cats) to end up with a stable representation of “cats.”
In analogy, the (re)activation of (novel combinations of) episodic experiences in the context of an unrelated or tangentially related thought or physical environment adds noise that might be useful for identifying regularities across instantiations. The data augmentation induced by mind wandering might involve noise consisting of partial activations of an experience (similar to occlusion), and/or disrupted spatiotemporal contiguity induced by the novel internal or external environment (similar to cropping). This theory suggests a potential role for mind wandering – and mental travel more generally – in facilitating abstraction through data augmentation. Adding to Gilead et al.'s theory, the ideas laid out here imply that the relationship between mental travel and abstraction is in fact mutualistic: abstraction facilitates mental travel, and mental travel facilitates abstraction. Taking into consideration the potentially mutualistic nature of this relationship is critical to understanding both mental travel and abstraction, as well as to understanding the function of the seemingly costly cognitive process of mind wandering.