ausblenden:
Schlagwörter:
-
Zusammenfassung:
The mammalian brain possesses multiple memory systems, including the striatum, the locus of procedural
memories underlying habits, the neocortex, the main site for long-term general perceptual and semantic
knowledge, and the hippocampus, the area largely responsible for storing and processing specific
autobiographical memories [1,2]. These three systems and their interactions have been extensively
studied, both empirically [3] and theoretically [4], but their collective roles in guiding optimal behavior
have only rarely been addressed [5]. The contribution of episodic memory is particularly mysterious.
Here, we develop a normative framework of control in the face of uncertainty, in which each plays a
precisely delineated part.
Uncertainty arises in control from inherent stochasticity of the underlying tasks, and ignorance of the
controller about the tasks, even if they are otherwise deterministic. Following others [6], we see the
neocortex as a learning system making efficient use of available information in such conditions. A key
aspect of such a system is that it represents not only single values of relevant variables, but also the
uncertainty surrounding those values, thereby at least approximately performing optimal statistical
inference [7]. However, we also observe that while it may be ideal – if at all possible – to keep account of
uncertainty for learning, ironically, it is precisely the same careful bookkeeping of uncertainty that renders
impractical planning sequences of actions using such a model. The more uncertainty there is to represent,
the less feasible it is to model all the consequences of taking an action by enumerating recursively all
possibilities that it entails in the future. This problem becomes exponentially harder as the time horizon of
the task broadens, making direct approximations, such as pruning or sampling, suffer from serious biases.
The striatal habit-based system has been suggested to offer a solution to this problem, but is only effective
in the limit of substantial samples [5].
We suggest that a different class of approximations, namely recalling episodic memories of specific
behavioral sequences that proved successful in the past, can be a powerful alternative to the other two
systems. The eventual reduction of ignorance-related subjective uncertainty and the steady accumulation
of sufficient information to license reliable habits, imply that the hippocampus should be particularly
important in the early stages of training on a task or exploring an environment.
Our results suggest normative accounts of the widely observed time-limited role of the hippocampus in
processing memories [2], and the apparently more semantic characteristics of distant memories [8]. This
offers a different perspective from the popular, but computationally challenging hypothesis that memories are consolidated out of the hippocampus and into the neocortex, or elsewhere [4].
