Coding and learning of behavioral sequences

Abstract

A major challenge to understanding behavior is how the nervous system allows the learning of behavioral sequences that can occur over arbitrary timescales, ranging from milliseconds up to seconds, using a fixed millisecond learning rule. This article describes some potential solutions, and then focuses on a study by Mehta et al. that could contribute towards solving this puzzle. They have discovered that an experience-dependent asymmetric shape of hippocampal receptive fields combined with oscillatory inhibition can serve to map behavioral sequences on a fixed timescale.Behavioral sequences of events span several orders of magnitude of timescales. A piano player is able to play rapid scales or repetitions of a single tone at a rate of 12 or more keys per second, probably using an internal sequence of motor commands that is even faster. However, when playing a slow tune, commands are executed ten or twenty times slower. Such different temporal sequences should ideally all be learned using a synaptic learning rule that operates on a fixed neuronal timescale of milliseconds [1 -3]. How then, is learning of behavioral sequences across arbitrary timescales possible? Several studies have recently addressed this issue from different points of view, and have suggested potential solutions.Coding sequences at millisecond resolution One potential solution of the problem could be that all sequences, fast or slow, are in fact stored and retrieved at a millisecond resolution. Synfire chains (SFC) [4] -dominantly feedforward multi-layered architecture (chains) in which spiking activity can propagate in a synchronous wave of neuronal firing (a pulse packet) from one layer of the chain to the next -provide a conceptual framework for such an approach. The chain of neuronal activity evolves on a millisecond scale and, hence, on the natural timescale of spike-time-dependent plasticity. Because all information is represented by a sequence of firing neuronal times, the SFC would automatically solve the problem of sequence learning using a fixed-timescale learning rule. The fastest behavioral sequence would be one in which each time step of the neuronal chain corresponds to one behavioral event. A slow behavioral sequence would then simply be represented by several neuronal time steps per behavioral event. The problem with such an approach is that the representation of slow sequences of events (e.g. five events in one second) is rather expensive, because it requires hundreds of transitions that need to be precisely stored in neuronal connections [4][5][6].A recent study suggested that a mechanism closely related to neuronal SFCs is implemented in the premotor area of song-birds [7]. Neurons projecting from the nucleus hyperstriatum ventralis pars caudalis (HVC) to the forebrain robust nucleus of the archistriatum (RA) are active in a tight temporal sequence or 'clock'. In contrast to the SFC concept, however, each neuron emits not a single spike but a burst of three to four spikes within ,10 ms. Learning a song could then occur in the...

Dynamical principles in neuroscience

“…A widely discussed hypothesis is that the specific circuitry of the IO, cerebellar cortex, and deep cerebellar nuclei called the slow loop ͑see Fig. 52͒ can serve as a dynamical working memory or as a neuronal clock with Ϸ100-ms cycle time which would make it easy to connect it to behavioral time scales ͑Kistler and de Zeeuw, 2002;Melamed et al, 2004͒. Temporal coordination and, in particular, synchronization of neural activity is a robust phenomenon, frequently observed across populations of neurons with diverse membrane properties and intrinsic frequencies. In the light of such diversity the question of how precise synchronization can be achieved in heterogeneous networks is critical.…”

Temporal Sequence Learning, Prediction, and Control: A Review of Different Models and Their Relation to Biological Mechanisms

“…Its is by now widely accepted that LTP (and LTD) are strongly involved in processes of learning and memory (Martin et al, 2000). On the other, at the moment there is no straight-forward way which would lead from an STDP model to, for example, a model of classical conditioning and only a few attempts exist in this direction (Mehta et al, 2002;Sato and Yamaguchi, 2003;Melamed et al, 2004). Maybe the serial compound stimulus representations, introduced in section 3.1.3, could be "recycled" in this context, but in itself such a representation seems to be a rather crude approximation of the neuronal interactions in the areas involved.…”

Phase Precession Through Synaptic Facilitation

“…Hippocampal phase precession might be important in sequence learning Mehta et al, 2002;Sato & Yamaguchi, 2003;Melamed et al, 2004;Jensen & Lisman, 2005;Lengyel et al, 2005). Our model thus provides a potential link between the physiology of the mossy fiber synapse and behavior (Lipp, Schwegler, Heimrich, & Driscoll, 1988).…”

“…Here we propose that synaptic facilitation, for example, at the hippocampal mossy fiber (mf) synapse, allows for generating a relational spike code. This code might be important for one-shot learning and episodic-like memory, that is, the association of events in a behavioral sequence that occur on a timescale of seconds (Skaggs, McNaughton, Wilson, & Barnes, 1996;Silva et al, 1996;Brun et al, 2002;Fortin, Agster, & Eichenbaum, 2002;Kesner, Gilbert, & Barua, 2002;Mehta, Lee, & Wilson, 2002;Sato & Yamaguchi, 2003;Melamed, Gerstner, Maass, Tsodyks, & Markram, 2004;Jensen & Lisman, 2005;Dragoi & Buzsáki, 2006).…”