Dynamic synapse: Harnessing the computing power of synaptic dynamics

Abstract

A major unresolved issues in neuroscience is the emergence of functional capability from underlying molecular and cellular mechanisms. The concept of a dynamic synapse is extended to provide a formalism to incorporate synaptic mechanisms into a general scheme of neural information processing: A synapse is composed of two distinct functional units, presynaptic terminals for transforming the sequence of action potentials into multiple sequences of discrete release events, and postsynaptic components for combining such synaptic signals. The complex interaction of various cellular and molecular processes in synapses can be concisely expressed and interpreted in these two synaptic terms. Learning involves modifying the synaptic dynamics such that each axon terminal performs proper transformation function.

Introduction

The brain derives its functional capacities from a vast number of underlying cellular and molecular mechanisms and their complex interactions. For instance, the existence of various dynamic processes in the axon terminal, including paired-pulse facilitation, augmentation, post-tetanus potentiation, depression, etc., has been well documented [4], [11], [21]. The process of neurotransmitter release is mediated by a complex molecular machinery [15], [19]. Furthermore, these molecular mechanisms themselves are dynamic, under constant regulation and modulation [9], [18]. Given such an immense complexity, how can we elucidate the role of synaptic dynamics in neural function? An even more tantalizing question is how can such regulation or modulation (e.g., the phosphorylation of say synaptotagmin, a protein in the axon terminal) be harnessed for learning and memory? Due to the complexity of the issue, the functional role of synaptic dynamics in neural information processing remains largely unexplored, except for a few studies (e.g. [16], [20]). In this article, we propose a formalism based on the concept of a dynamic synapse [6] for incorporating synaptic mechanisms into the general scheme of information processing in the brain.

The dynamic synapse proposed by Liaw and Berger (1996) embodies three major concepts. First, an axon terminal transforms a sequence of action potentials into another sequence of discrete release events. This is a functional consequence of the existence of multiple excitatory and inhibitory mechanisms with different time constants in the terminal. Moreover, once released, the rise and fall of the concentration of neurotransmitter in the synaptic cleft is very rapid, with a decay time constant of 1–2 ms [2]. Thus, a release event can be regarded as a “spike”, much the same way as an action potential is referred to. Second, an array of different sequences of release events can be generated by the axon terminals of a single neuron. This is a result of variations in the synaptic mechanisms across different presynaptic terminals [5], [17] In other words, each axon terminal acts as an independent communication channel such that information is coded in the spatio-temporal patterns of release events, instead of a single sequence of action potentials. Third, during learning, presynaptic mechanisms are modified such that each axon terminal performs proper pattern transformation function while changes in postsynaptic mechanisms determine how the synaptic signals are to be combined. The computational capability of synaptic dynamics has been demonstrated by achieving speaker-independent word recognition from raw waveforms by a small and simple neural network incorporating dynamic synapses [6], [7]. The dynamic synapse neural network exhibits a high degree of robustness against noise [8].

More recently, there is an increased interest in the functional role of synaptic dynamics [10], [12], [13], [14], [16]. For example, concepts similar to the first and second points of the aforementioned dynamic synapse has been proposed in [10], [14]. Maass and Zador (1998) incorporated the processes of facilitation and depression in their “dynamic stochastic synapse” as a basic computational unit in neural networks. Markram and colleagues (1998) also include facilitation and depression in their notion of “utilization of synaptic efficacy” and the first two points of dynamic synapse were clearly echoed in their proposed concept of “differential signaling” as a basis for information processing within networks of neurons.

Here we describe (1) how various synaptic mechanisms can be incorporated into networks of neurons via dynamic synapses as a way of understanding their functional role in neural information processing, and (2) how the synaptic dynamics can be modified as a basis of learning and memory.