Agent: Research community—Systems Neuroscience

Situation:

Systems consolidation theory (SCT, Squire & Alvarez, 1995) and complementary learning systems theory (CLS, McClelland et al., 1995) are prominent theories in the field of systems neuroscience. Roughly, CLS claims that the mammalian brain has two complementary learning systems: one, located in the neocortex, is the basis for the gradual acquisition of structured knowledge about an animal’s environment, while the other, located in the hippocampal formation, supports rapid learning of individual experiences. SCT maintains that the offline replay of hippocampally stored memories supports the transfer of new information from those experiences into broadly distributed circuits across the neocortex.

These theories can account for a wide range of experimental phenomena. For example, they are used to explain the effects of anterograde amnesia following hippocampal lesions (Scoville & Milner, 1957; Squire, 1992) compared to the apparent lack of memory effects following localized cortical lesions (Lashley, 1950), the presence and spatiotemporal coupling of replay and reactivation in hippocampal and cortical circuits during sleep (Girardeau & Lopes-dos-Santos, 2021; Ji & Wilson, 2007; Wilson & McNaughton, 1994), the effects of targeted perturbations during sleep and learning via electrical or optical methods (Girardeau et al., 2009; Maingret et al., 2016), and the spatial localization and behavioral dependence on memory-associated “engram” cells after learning (Kitamura et al., 2017). Further, CLS/SCT has been informative for a range of applied and extrascientific problems—for example, they have inspired the development of artificial neural network architectures (Mnih et al., 2015) and explain patterns of memory deficits in epilepsy (Gelinas et al., 2016) and Alzheimer’s disease (Zhen et al., 2021), which has lead to potential therapeutic targets (Lee et al., 2020).

However, there are a number of experimental observations which are not accounted for by the theories. For example, it has been observed that the ability to recall information learned during some experiences remains dependent on the hippocampus for the entire lifetime of an animal (Gilboa & Moscovitch,  2021). This is a shortcoming of the systems neuroscience’s body of knowledge, as it indicates that the theory will not be able to effectively inform, e.g., further experiments (how to design an experiment in a new memory paradigm when the predominant theory cannot predict whether the memory will be hippocampal or cortical-dependent?) or potential external agents who wish to use the theory (how to develop treatments to alleviate epilepsy-related memory symptoms when your theory on which your treatment is based is not reliable?).

Constraints:

The primary constraint is that a solution should answer the research question, by providing an explanation for the observations that systems consolidation only applies to a subset of hippocampal memories. The question is framed as a “why?” question, for which a frequent approach in the field is to provide a normative explanation which demonstrates that some state of affairs is optimal for the solution of some task under neurally plausible conditions (Levenstein et al., 2023). A solution should thus consist of a statement of the “task” being performed by systems consolidation and a demonstration that the observations (selective memory consolidation) are beneficial for its solution. The task, as well as the conditions under which it is optimized (which are also frequently called “constraints”), should be supportable by existing knowledge in the field. In addition, a solution would ideally not disrupt the ability of the field’s body of knowledge to account for other, already solved, problems using SCT/CLS.

Proposed Solution:

The authors propose a modification of systems consolidation theory (“generalization-optimized systems consolidation”), according to which hippocampal memories are only consolidated when it aids generalization. In support of this solution, the authors introduce a new neural network formalization of systems consolidation which reveals an overlooked tension in SCT: unregulated neocortical memory transfer can cause overfitting and harm generalization in unpredictable environments. Thus, the observation of selective consolidation can be explained by a postulate that memories which remain hippocampal-dependent are not generalizable.