Trends in Neurosciences
Hippocampal sequence-encoding driven by a cortical multi-item working memory buffer
Introduction
It is important to identify the synaptic and network mechanisms that support the learning and recall of sequences [1]. To learn a sequence, synaptic linkages must be strengthened between cells representing sequential items in the sequence. These linkages make possible the subsequent recall of the sequence. For instance, the presentation of the cue A would lead to the activation of the cells encoding A. Then, through the strengthened linkages [2], the firing of A cells would lead to the firing of B cells and, through similar chaining, evoke the rest of the sequence. The forms of LTP so far discovered have a time window for induction of ∼100 ms 3, 4, 5 and cannot therefore strengthen connections between cells that fire with greater temporal separation. Because humans can learn sequences with item separation of seconds, some other mechanism must be involved. Here, we argue that the additional mechanism could be a cortical buffer capable of holding multiple items in a short-term memory buffer (used here synonymously with working memory). We argue that such a buffer can represent items in a way that preserves their order and can drive the encoding of realistic sequences within the hippocampal long-term memory system (Figure 1).
Section snippets
Hippocampal involvement in sequence memory
Before considering the role of buffers in sequence encoding, we briefly review the evidence that implicates the hippocampus in the long-term storage of sequences. Various behavioral studies have shown that hippocampal lesions interfere with behaviors that require sequence learning while leaving simple forms of item recognition intact 6, 7. Electrophysiological studies have also provided evidence for hippocampal involvement. For instance, the sequences of places the animal experienced while
Evidence of a short-term memory buffer
Behavioral analysis of human memory studies has revealed important properties of short-term memory. In a recall experiment, a list of n items is presented and the subject is asked to repeat the list in order. For n<7, recall is nearly perfect, but for n>7 recall falls dramatically [16]. This led to the idea that the short-term memory buffer can store ∼7 items (chunks) and drive the incorporation of information into long-term stores. Models of this kind account for a wide range of measured
A physiologically based model of a neocortical multi-item buffer: the role of theta and gamma oscillations
Figure 3a shows the model we have put forward for multi-item working memory 40, 41. The idea is that the same network can store multiple items using a temporal multiplexing mechanism that is clocked by the theta and gamma oscillations known to exist in these structures [33]. According to this model, the group of cells that represents a single (chunked) item fires on each theta cycle, but only in a given gamma subcycle. Sequential items are active in sequential gamma subcycles, thereby encoding
In learning mode, a multi-item buffer can drive hippocampal sequence learning by standard LTP
In the multi-item theta–gamma buffer, items presented over many seconds will be active in sequential gamma cycles – that is, with a temporal separation of ∼30 ms (Figure 3a), well within the 100 ms time-window of hippocampal LTP 4, 5. When it is desirable to encode a sequence into hippocampal long-term memory, the output of the multi-item buffer will be funneled into the hippocampus (Figure 1a). Simulations demonstrate that as multiple items fire in the buffer, they can become reliably encoded in
Closing Remarks
Recent work has suggested a process of sequence encoding that is different from that proposed here [11]. As outlined in Figure 1c, we view the phase precession as a result of sequence encoding. By contrast, other groups have argued that the importance of the phase precession might instead be to allow sequence encoding 11, 51. According to the model of Mehta et al. [11], the connections of the feedforward CA3 synapses onto CA1 are modified by LTP to produce a ramp of firing as the rat passes
Acknowledgements
This research was (in part) supported in the framework of the Netherlands Organization for Scientific Research (NWO) Innovative Research Incentive Schemes, with financial aid from the NWO and grant 1 R01 NS50944-01 as part of the NSF/NIH Collaborative Research in Computational Neuroscience Program.
References (57)
Coding and learning of behavioral sequences
Trends Neurosci.
(2004)- et al.
Memory of sequential experience in the hippocampus during slow wave sleep
Neuron
(2002) - et al.
Human memory: a proposed system and its control processes
The episodic buffer: a new component of working memory?
Trends Cogn. Sci.
(2000)Human hippocampal formation EEG desynchronizes during attentiveness and movement
Electroencephalogr. Clin. Neurophysiol.
(1978)The spectral properties of hippocampal EEG related to behaviour in man
Electroencephalogr. Clin. Neurophysiol.
(1980)Trajectory encoding in the hippocampus and entorhinal cortex
Neuron
(2000)- et al.
Prospective and retrospective memory coding in the hippocampus
Neuron
(2003) Place cell firing shows an inertia-like process
Neurocomputing
(2000)Modeling goal-directed spatial navigation in the rat based on physiological data from the hippocampal formation
Neural Netw.
(2003)