We commend Bastin et al. on developing an integrative dual-process model of recognition memory that considers the role of distinct brain regions in representing information, and in making attributions about experience-dependent changes to these representations, in memory decisions. In our view, such an integration has been missing in the extant cognitive neuroscience literature, which has typically focused exclusively either on representations or on cognitive processes when characterizing the role of different structures (e.g., Bussey & Saksida Reference Bussey and Saksida2007 versus Brown & Aggleton Reference Brown and Aggleton2001). Past accounts of recognition memory that have made reference to attribution have discussed it in relation to fluency, with attribution of fluency to prior experience being at the core of familiarity-based judgments (Dew & Cabeza Reference Dew and Cabeza2013; Jacoby et al. Reference Jacoby, Kelley, Dywan, Roediger and Craik1989). In the current model, the authors take a similar stance when specifying the role of perirhinal cortex (PrC) and anterolateral entorhinal cortex in providing fluency signals. As the authors acknowledge, this fluency account contrasts, at least on the surface, with another dominant account of familiarity assessment that focuses on global-matching computations, which have also been linked to PrC (LaRocque et al. Reference LaRocque, Smith, Carr, Witthoft, Grill-Spector and Wagner2013; Norman Reference Norman2010).
We would like to point out that global matching and fluency accounts of familiarity may not be mutually exclusive. In the integrative model proposed here, fluency can arise from repetition (i.e., prior exposure) of perceptual or conceptual features at different levels of a representational hierarchy, with PrC being sensitive to repetition at the entity level where features are highly conjunctive and can differentiate between different exemplars of objects with high feature overlap. Critically, feature overlap also plays a key role in global matching and has been linked to behavioral evidence, such as false alarm rates to lures similar to targets, in recognition-memory judgments (Montefinese et al. Reference Montefinese, Zannino and Ambrosini2015). In the influential MINERVA 2 model (Hintzman Reference Hintzman1984) of global matching in recognition memory, a retrieval cue induces an echo whose intensity is directly based on a scalar measure of feature overlap between the cue and all stored memory traces. Fluency may be a signal that simply reflects this intensity measurement.
Global matching and fluency can also be linked to a common neural phenomenon in terms of changes to representations that occur with repeated exposures: namely, repetition suppression. Repetition suppression is well documented in the perirhinal cortex (Suzuki & Naya Reference Suzuki and Naya2014) and has been suggested to reflect a fluency signal that can inform decisions on a variety of tasks, including but not limited to familiarity-based memory judgments (Dew & Cabeza Reference Dew and Cabeza2013). Although the functional significance and underlying mechanisms of repetition suppression in neural recordings remain contentious (Barron et al. Reference Barron, Garvert and Behrens2016; Grill-Spector et al. Reference Grill-Spector, Henson and Martin2006), at least one of the proposed mechanisms, “sharpening,” can support both computations of global matching and fluency signaling. In a sharpening account, neural representations of a stimulus become sparser over repetitions, as neurons that initially responded weakly to a stimulus gradually “drop out.” In the complementary learning system neural network model (Norman & O'Reilly Reference Norman and O'Reilly2003; see also Norman Reference Norman2010), such sharpening is the result of a competitive Hebbian learning process between neurons in neocortical regions; it is linked to global matching by virtue of stimuli with high degree of feature overlap also being represented with overlapping neural patterns. Inasmuch as repetition suppression in single cell recordings and in fMRI (functional magnetic resonance imaging) BOLD signals is not limited to the PrC, and has also been shown to occur, for example, in other ventral visual pathway regions (Barron et al. Reference Barron, Garvert and Behrens2016), wide-spread repetition suppression effects are consistent with the proposal in the present integrative memory model that fluency signals can arise at multiple levels.
Considering global-matching computations (and their link to fluency) may also be of value when trying to understand the mechanisms that underlie the attribution process in recognition memory as proposed in the integrative memory model. It is our impression that this attribution system is currently less well specified, and supported by less empirical evidence overall, than the proposed representation system. In the integrative memory model, the attribution system interprets changes in representations toward the goal of making overt memory decisions. A promising account that may help to elaborate on how attribution processes lead to memory judgments is provided by the drift-diffusion model (Ratcliff Reference Ratcliff1978; see also Ratcliff et al. Reference Ratcliff, Smith, Brown and McKoon2016b). This model addresses the temporal unfolding of memory retrieval and treats the comparison of feature overlap between cues and stored traces during this retrieval process as accumulating noisy evidence. Because all memory traces are compared in parallel, these computations can be understood as global matching, with fluency reflecting the combined speed of these parallel accumulation streams.
An emerging body of evidence from functional neuroimaging and other recording techniques points to a role for lateral parietal cortex in evidence accumulation during decision making, including but not limited to memory judgments (Wagner et al. Reference Wagner, Shannon, Kahn and Buckner2005). Some studies have even identified specific neurons in the lateral intraparietal sulcus whose activity profile can be interpreted as evidence accumulation (Shadlen & Newsome Reference Shadlen and Newsome2001). Against this background, the specification of structures involved in memory attribution in the integrative memory model may require expansion beyond prefrontal cortex, and additional emphasis on lateral parietal cortex as a key player. At present, the latter structure is primarily concerned with attentional mechanisms in this model. There is some evidence to suggest, however, that attention effects observed in the lateral parietal lobe are at least in part spatially distinct from memory effects (Hutchinson et al. Reference Hutchinson, Uncapher and Wagner2009; Reference Hutchinson, Uncapher, Weiner, Bressler, Silver, Preston and Wagner2014). Therefore, exclusive reference to attentional mechanisms may not fully capture its role in attribution processes as part of the decision making just described.
We commend Bastin et al. on developing an integrative dual-process model of recognition memory that considers the role of distinct brain regions in representing information, and in making attributions about experience-dependent changes to these representations, in memory decisions. In our view, such an integration has been missing in the extant cognitive neuroscience literature, which has typically focused exclusively either on representations or on cognitive processes when characterizing the role of different structures (e.g., Bussey & Saksida Reference Bussey and Saksida2007 versus Brown & Aggleton Reference Brown and Aggleton2001). Past accounts of recognition memory that have made reference to attribution have discussed it in relation to fluency, with attribution of fluency to prior experience being at the core of familiarity-based judgments (Dew & Cabeza Reference Dew and Cabeza2013; Jacoby et al. Reference Jacoby, Kelley, Dywan, Roediger and Craik1989). In the current model, the authors take a similar stance when specifying the role of perirhinal cortex (PrC) and anterolateral entorhinal cortex in providing fluency signals. As the authors acknowledge, this fluency account contrasts, at least on the surface, with another dominant account of familiarity assessment that focuses on global-matching computations, which have also been linked to PrC (LaRocque et al. Reference LaRocque, Smith, Carr, Witthoft, Grill-Spector and Wagner2013; Norman Reference Norman2010).
We would like to point out that global matching and fluency accounts of familiarity may not be mutually exclusive. In the integrative model proposed here, fluency can arise from repetition (i.e., prior exposure) of perceptual or conceptual features at different levels of a representational hierarchy, with PrC being sensitive to repetition at the entity level where features are highly conjunctive and can differentiate between different exemplars of objects with high feature overlap. Critically, feature overlap also plays a key role in global matching and has been linked to behavioral evidence, such as false alarm rates to lures similar to targets, in recognition-memory judgments (Montefinese et al. Reference Montefinese, Zannino and Ambrosini2015). In the influential MINERVA 2 model (Hintzman Reference Hintzman1984) of global matching in recognition memory, a retrieval cue induces an echo whose intensity is directly based on a scalar measure of feature overlap between the cue and all stored memory traces. Fluency may be a signal that simply reflects this intensity measurement.
Global matching and fluency can also be linked to a common neural phenomenon in terms of changes to representations that occur with repeated exposures: namely, repetition suppression. Repetition suppression is well documented in the perirhinal cortex (Suzuki & Naya Reference Suzuki and Naya2014) and has been suggested to reflect a fluency signal that can inform decisions on a variety of tasks, including but not limited to familiarity-based memory judgments (Dew & Cabeza Reference Dew and Cabeza2013). Although the functional significance and underlying mechanisms of repetition suppression in neural recordings remain contentious (Barron et al. Reference Barron, Garvert and Behrens2016; Grill-Spector et al. Reference Grill-Spector, Henson and Martin2006), at least one of the proposed mechanisms, “sharpening,” can support both computations of global matching and fluency signaling. In a sharpening account, neural representations of a stimulus become sparser over repetitions, as neurons that initially responded weakly to a stimulus gradually “drop out.” In the complementary learning system neural network model (Norman & O'Reilly Reference Norman and O'Reilly2003; see also Norman Reference Norman2010), such sharpening is the result of a competitive Hebbian learning process between neurons in neocortical regions; it is linked to global matching by virtue of stimuli with high degree of feature overlap also being represented with overlapping neural patterns. Inasmuch as repetition suppression in single cell recordings and in fMRI (functional magnetic resonance imaging) BOLD signals is not limited to the PrC, and has also been shown to occur, for example, in other ventral visual pathway regions (Barron et al. Reference Barron, Garvert and Behrens2016), wide-spread repetition suppression effects are consistent with the proposal in the present integrative memory model that fluency signals can arise at multiple levels.
Considering global-matching computations (and their link to fluency) may also be of value when trying to understand the mechanisms that underlie the attribution process in recognition memory as proposed in the integrative memory model. It is our impression that this attribution system is currently less well specified, and supported by less empirical evidence overall, than the proposed representation system. In the integrative memory model, the attribution system interprets changes in representations toward the goal of making overt memory decisions. A promising account that may help to elaborate on how attribution processes lead to memory judgments is provided by the drift-diffusion model (Ratcliff Reference Ratcliff1978; see also Ratcliff et al. Reference Ratcliff, Smith, Brown and McKoon2016b). This model addresses the temporal unfolding of memory retrieval and treats the comparison of feature overlap between cues and stored traces during this retrieval process as accumulating noisy evidence. Because all memory traces are compared in parallel, these computations can be understood as global matching, with fluency reflecting the combined speed of these parallel accumulation streams.
An emerging body of evidence from functional neuroimaging and other recording techniques points to a role for lateral parietal cortex in evidence accumulation during decision making, including but not limited to memory judgments (Wagner et al. Reference Wagner, Shannon, Kahn and Buckner2005). Some studies have even identified specific neurons in the lateral intraparietal sulcus whose activity profile can be interpreted as evidence accumulation (Shadlen & Newsome Reference Shadlen and Newsome2001). Against this background, the specification of structures involved in memory attribution in the integrative memory model may require expansion beyond prefrontal cortex, and additional emphasis on lateral parietal cortex as a key player. At present, the latter structure is primarily concerned with attentional mechanisms in this model. There is some evidence to suggest, however, that attention effects observed in the lateral parietal lobe are at least in part spatially distinct from memory effects (Hutchinson et al. Reference Hutchinson, Uncapher and Wagner2009; Reference Hutchinson, Uncapher, Weiner, Bressler, Silver, Preston and Wagner2014). Therefore, exclusive reference to attentional mechanisms may not fully capture its role in attribution processes as part of the decision making just described.