A central aspect of the target article is the neocortical junction that is encoded in the hippocampus as an index. Hyperassociative rapid eye movement (REM) dreams could be retained to constitute the hippocampal index. During memory retrieval, elicitation of the appropriate index from the hippocampus by internal or external stimuli enables the subsequent retrieval of cortical content using the index. In the elaboration of this proposal, section 5.1 poses two questions: “How does a non-conscious hyperassociative index trigger conscious veridical episodic (i.e., situated in time and place) output? And how is the conscious output constrained to only the relevant episodic memory?” (para. 1).
Llewellyn then provides a descriptive account of how the hippocampus and neocortex interact to achieve this. However, the functional implementation of such a mechanism is not specified. We can consider an analysis of the functional description of this REM-based hippocampal indexing, and the subsequent memory retrieval, in the concrete context of implemented recurrent network models of cortex that we traditionally have employed in sensorimotor sequence learning (Dominey Reference Dominey1995) and in language processing (Dominey & Ramus Reference Dominey and Ramus2000; Hinaut & Dominey Reference Hinaut and Dominey2013). In this context, static patterns or snapshots of cortical activity that can be stored as hyperassociative indices in hippocampus can subsequently be retrieved and reinjected into the neocortex to enable neocortex to then unfold the corresponding sequence, thus implementing an index-based sequence memory storage and retrieval capability. This is potentially interesting because it provides a form of validation (if successful) of a mechanism that is provided only as a possible solution in the target article. Such an analysis suggests that the index is not a cortical locus or set of loci, but rather a snapshot of the cortical state at that time which can be used, as stated in the target text, as a cue for retrieval in an autoassociative memory. Before addressing the two questions posed earlier, we can first consider these issues: What is the nature of the neocortical junction, and how can it be used in memory retrieval? This first poses the question of what the nature of a conscious memory is.
Memories will tend to implicate the semantic system, which has been demonstrated to encompass a broadly extended network of distributed cortical areas (Binder & Desai Reference Binder and Desai2011; Binder et al. Reference Binder, Desai, Graves and Conant2009). In this context, the activation necessary to invoke a memory could involve a fairly massive activation of a large distribution of the neocortex. It has been suggested that the hippocampus integrates distributed cortical activity, fusing this coactivation into a memory trace, and that over time the cortex can become independent of hippocampus, with prefrontal cortex taking over the role of integration for more mature memories (Frankland & Bontempi Reference Frankland and Bontempi2005). This suggests the hippocampus would be able to re-instantiate a prior state of cortical activation. Once this state of activation is instantiated, the cortex would then play out the corresponding memory.
We have modeled cortex as a dynamic system of leaky integrator neurons with local recurrent connections (Dominey Reference Dominey1995; Hinaut & Dominey Reference Hinaut and Dominey2013). Such networks have interesting dynamics. In particular, the internal state follows a trajectory such that if the system is put into a state along an existing trajectory, then the system will tend to follow that trajectory from the given state as a point of departure. Based on this property, the state of activation of cortex could be stored as an index by the hippocampus and then reinjected into the cortex. In such conditions, the cortex would then continue in the appropriate trajectory from that point onward, thus “replaying” the corresponding dynamic memory trace. Importantly, such systems display some robustness to noise, but also a form of degraded behavior: if the injected pattern deviates sufficiently from the intended pattern, then the resulting trajectory will deviate from the intended trajectory. This implies that the pattern of activity that is played into the cortex from the hippocampus should be as accurate and complete as possible.
In other words, if a specific memory is to be recalled, then it should be indexed in the most specific manner possible. This suggests, as indicated by Llewellyn, that hippocampus keeps an index of multiple loci that can be used in episodic memory retrieval. If sufficient loci are activated, then the cortex will enter into a state from which a dynamic trajectory will then unfold.
This trajectory can be considered to correspond to the narrativization of experience into a linear sequence. The question that remains, with respect to junctions, is if a trajectory proceeds through a junction, how can the system ensure that it does not deviate onto a different trajectory that traverses that junction? That is, how is the system constrained to recall only the intended or relevant episodic memory? From the perspective of the dynamic systems and recurrent network models that we manipulate, the more that the pattern of cortical activity – entrained by the hippocampal index – is complete and corresponds to the memory to be recalled, the more that the resulting trajectory of cortical activity will correspond to the unfolding of the corresponding episodic memory.
This comment thus advocates the characterization of the cortex as a dynamic system, with state trajectories that can be “replayed” by putting cortex into a past state via hippocampal inputs. This provides a mechanism that is consistent with Llewellyn's proposal and provides potential responses to the questions: “How does a non-conscious hyperassociative index trigger conscious veridical episodic (i.e., situated in time and place) output? And how is the conscious output constrained to only the relevant episodic memory?”
You have
Access
A central aspect of the target article is the neocortical junction that is encoded in the hippocampus as an index. Hyperassociative rapid eye movement (REM) dreams could be retained to constitute the hippocampal index. During memory retrieval, elicitation of the appropriate index from the hippocampus by internal or external stimuli enables the subsequent retrieval of cortical content using the index. In the elaboration of this proposal, section 5.1 poses two questions: “How does a non-conscious hyperassociative index trigger conscious veridical episodic (i.e., situated in time and place) output? And how is the conscious output constrained to only the relevant episodic memory?” (para. 1).
Llewellyn then provides a descriptive account of how the hippocampus and neocortex interact to achieve this. However, the functional implementation of such a mechanism is not specified. We can consider an analysis of the functional description of this REM-based hippocampal indexing, and the subsequent memory retrieval, in the concrete context of implemented recurrent network models of cortex that we traditionally have employed in sensorimotor sequence learning (Dominey Reference Dominey1995) and in language processing (Dominey & Ramus Reference Dominey and Ramus2000; Hinaut & Dominey Reference Hinaut and Dominey2013). In this context, static patterns or snapshots of cortical activity that can be stored as hyperassociative indices in hippocampus can subsequently be retrieved and reinjected into the neocortex to enable neocortex to then unfold the corresponding sequence, thus implementing an index-based sequence memory storage and retrieval capability. This is potentially interesting because it provides a form of validation (if successful) of a mechanism that is provided only as a possible solution in the target article. Such an analysis suggests that the index is not a cortical locus or set of loci, but rather a snapshot of the cortical state at that time which can be used, as stated in the target text, as a cue for retrieval in an autoassociative memory. Before addressing the two questions posed earlier, we can first consider these issues: What is the nature of the neocortical junction, and how can it be used in memory retrieval? This first poses the question of what the nature of a conscious memory is.
Memories will tend to implicate the semantic system, which has been demonstrated to encompass a broadly extended network of distributed cortical areas (Binder & Desai Reference Binder and Desai2011; Binder et al. Reference Binder, Desai, Graves and Conant2009). In this context, the activation necessary to invoke a memory could involve a fairly massive activation of a large distribution of the neocortex. It has been suggested that the hippocampus integrates distributed cortical activity, fusing this coactivation into a memory trace, and that over time the cortex can become independent of hippocampus, with prefrontal cortex taking over the role of integration for more mature memories (Frankland & Bontempi Reference Frankland and Bontempi2005). This suggests the hippocampus would be able to re-instantiate a prior state of cortical activation. Once this state of activation is instantiated, the cortex would then play out the corresponding memory.
We have modeled cortex as a dynamic system of leaky integrator neurons with local recurrent connections (Dominey Reference Dominey1995; Hinaut & Dominey Reference Hinaut and Dominey2013). Such networks have interesting dynamics. In particular, the internal state follows a trajectory such that if the system is put into a state along an existing trajectory, then the system will tend to follow that trajectory from the given state as a point of departure. Based on this property, the state of activation of cortex could be stored as an index by the hippocampus and then reinjected into the cortex. In such conditions, the cortex would then continue in the appropriate trajectory from that point onward, thus “replaying” the corresponding dynamic memory trace. Importantly, such systems display some robustness to noise, but also a form of degraded behavior: if the injected pattern deviates sufficiently from the intended pattern, then the resulting trajectory will deviate from the intended trajectory. This implies that the pattern of activity that is played into the cortex from the hippocampus should be as accurate and complete as possible.
In other words, if a specific memory is to be recalled, then it should be indexed in the most specific manner possible. This suggests, as indicated by Llewellyn, that hippocampus keeps an index of multiple loci that can be used in episodic memory retrieval. If sufficient loci are activated, then the cortex will enter into a state from which a dynamic trajectory will then unfold.
This trajectory can be considered to correspond to the narrativization of experience into a linear sequence. The question that remains, with respect to junctions, is if a trajectory proceeds through a junction, how can the system ensure that it does not deviate onto a different trajectory that traverses that junction? That is, how is the system constrained to recall only the intended or relevant episodic memory? From the perspective of the dynamic systems and recurrent network models that we manipulate, the more that the pattern of cortical activity – entrained by the hippocampal index – is complete and corresponds to the memory to be recalled, the more that the resulting trajectory of cortical activity will correspond to the unfolding of the corresponding episodic memory.
This comment thus advocates the characterization of the cortex as a dynamic system, with state trajectories that can be “replayed” by putting cortex into a past state via hippocampal inputs. This provides a mechanism that is consistent with Llewellyn's proposal and provides potential responses to the questions: “How does a non-conscious hyperassociative index trigger conscious veridical episodic (i.e., situated in time and place) output? And how is the conscious output constrained to only the relevant episodic memory?”