Elsevier

Neurocomputing

Volume 73, Issues 1–3, December 2009, Pages 381-388
Neurocomputing

Synchronization enhances synaptic efficacy through spike timing-dependent plasticity in the olfactory system

https://doi.org/10.1016/j.neucom.2009.08.003Get rights and content

Abstract

Synaptic modifications are measured in biological experiments with respect to spike timings. Spike timing-dependent plasticity is the latest development in refinements of Hebbian learning. We have applied additive and multiplicative STDP synaptic learning rules to a biologically inspired olfactory network. The olfactory system recognizes odorant patterns by synchronization of mitral cells. Synchronization enhances synaptic connections between mitral cells and cortical cells. Both STDP rules exhibit unimodal weight distributions which is biologically realistic. As a result, cortical cells respond with a wider range of variability and higher firing frequency. This property has potential for the improvement of artificial odor recognition through ongoing selection of mitral cells.

Introduction

The brain is the center of intelligence, necessary for learning new concepts and dealing with uncertainty. Plasticity among neuronal connections in neural networks is believed to be the major mechanism that underlies intelligence [1]. Information in neural networks is encoded in the form of spike trains. Spike trains spread from one neuron to another through synapses and finally activate specific neurons in the cortex. The spatial position, spiking time and firing frequency of spike trains are studied by researchers to reveal the coding by the brain. However, it is still an open problem how information is represented in the brain using spiking trains.

We experience learning new concepts through reenforcement of practice. For instance, we remember a telephone number by repeating it several times. Hebbian learning was introduced to neural networks in this scenario by Hebb in 1949. Hebb proposed that “when one cell repeatedly assists in firing another, the axon of the first cell develops synaptic knobs (or enlarges them if they already exist) in contact with the soma of the second cell” [1]. Hebbian theory was confirmed by experimental discoveries of long-term potentiation (LTP) in the rabbit hippocampus in 1966 [2]. LTP describes long-term enhancement of a synapse between two neurons, which are stimulated at the same time. Although there is no sufficient evidence, long-term depression (LTD) is also postulated in neural networks to balance enhancement caused by LTP. A lot of correlation learning rules inspired by Hebbian theory have proved to be successful in specific neural networks [3].

In 1998, spike timing was found to be critical to synaptic modification [4]. Synaptic changes depending on the timing of both pre- and postsynaptic spike trains are named spike timing-dependent plasticity (STDP). Feedback of postsynaptic neurons can reflect the synaptic weights locally with global knowledge of the network. There are some variations of STDP and details of differences are explained in Section 3.

STDP has been discovered to facilitate synchronization in the olfactory neural networks of locusts [5], [6]. On the contrary, we have explored the features of STDP in a closed network with spontaneous synchronization [7]. Synchronization is evoked by external inputs in a recurrent network with frequency-dependent synapses [8]. For invertebrate, Finelli et al. have proposed that STDP regulates and selects specific output of projection neurons to form sparsely firing of Kenyon cells in the mushroom body [6].

In the biological olfactory system, odorants are bound to their specific receptors residing in the capillary network around sensory neurons. One neuron can only expresses one kind of receptor and neurons with the same receptors are classified as the same type of sensory neurons. These sensory neurons play an important role in the transition from chemical detection to electrical signals, which initiates olfactory perceptions.

The same type of sensory neurons are assembled at the same location in the olfactory bulb, named glomerulus. The number of glomeruli is close to the types of receptors. Activity patterns of firing rates are repeatable by odorant exposures. These patterns are believed to be the representations of odors in the olfactory bulb.

The olfactory bulb connects to the cortex via mitral/tufted cells. The processing in the cortex is complicated and recent research has to focuses on a specific small region of the cortex (mushroom body for invertebrates) [9]. For simplicity, a cortical cell is used to represent the activity of one odor in the cortex in our model. Further details of the olfactory system can be found in our early work [10].

In this paper, we study how synchronization affects STDP in a feed-forward network and the potential function of STDP for synchronization in an olfactory system we have built [10]. Features of two types of STDPs, weight-dependent and weight-independent STDP, respectively, are compared to validate the advantages and disadvantages of both models. After the comparison, we apply two kinds of STDP to a previous model of the olfactory neural network and examine the effects of both STDP within the olfactory system. In the olfactory model, selective mitral cells tend to fire synchronously with each other when the target odor presents. This synchronization will therefore modify the connection strength between mitral cells and cortical cells. Synchronization is found to increase the response variability of cortical neurons and thus makes the model more realistic to neurons in the waking brain [11]. In the simulation, both STDPs potentiate connection strengths when synchronization appears and encourages more firing in the cortical cells. While non-synchronized mitral cells depress synaptic weights or even eliminate the connections, the weight becomes zero. This selection can affect the olfactory system in the evolution of mitral cell selection [12], [13], [14].

Section snippets

Olfactory model

The topology of the system we constructed refers to biological olfactory systems of vertebrates like humans. The olfactory system of invertebrates like honeybees and locusts is sharing a similar structure of vertebrates. As indicated in Buck and Axel's paper, there are about 350 types of sensory receptor genes for humans and one sensory neuron can only express one species of gene [15]. All the sensory neurons expressing the same gene are dedicated to one glomerulus. The transduction between

Spike timing-dependent plasticity

Bi and Poo showed that spike timing of pre- and postsynaptic neurons is relative to the efficacy of their synapse in cultured rat hippocampus [4]. A critical window of synaptic modification was defined according to the time difference between pre- and postsynaptic spikes. Song, Miller and Abbott introduced a spike timing-dependent plasticity (STDP) model to simulate Bi and Poo's experimental results [12]. The STDP rule is defined as F(Δt)=A+exp(Δt/τ+)if Δt<0,-A-exp(-Δt/τ-)if Δt0,where Δt is

Synchronous effects on synaptic connections and efferent neurons

We have constructed an olfactory model using Borisyuk's integrate and fire neurons [25]. Descriptions of Borisyuk's neuron model are not supplied in this paper, for full details the reader may refer to Borisyuk's paper [25] and our previous work [10], [26].

With an external odor stimulus, a selective group of mitral cells fire synchronously. This group of mitral cells converge to a cortical cell. The integration window of cortical cells is about 30 ms and coincidence detection is used in our

Summary

Spike timing-dependent plasticity is a temporal synaptic learning rule, in which synaptic modification is updated according to the spike times of presynaptic and postsynaptic spike trains. Our biologically inspired olfactory network model was presented at the IJCNN conference 2007 [10]. The neural network is simulated using Borisyuk's integrate and fire neuron model, which is a sophisticated neuron model containing decay, delay, refractory period and noise.

In the olfactory bulb of the system,

Xiaobin Lin is a Ph.D. candidate in the Department of Computer Science, Heriot-Watt University, Edinburgh, United Kingdom. He obtained a M.Sc. degree in Information and Intelligence from the Electronics and Electrical Engineering Department, University of Liverpool, United Kingdom, in 2005 and a B.S. degree in Computer Science from the Computer Science Department, Shenzhen University, Shenzhen, China, in 2004. His research interests include neural network topology and plasticity in the brain,

References (30)

  • S. Cassenaer et al.

    Hebbian STDP in mushroom bodies facilitates the synchronous flow of olfactory information in locusts

    Nature

    (2007)
  • L. Finelli, S. Haney, M. Bazhenov, M. Stopfer, T. Sejnowski, Synaptic learning rules and sparse coding in a model...
  • X. Shen et al.

    Oscillations and spiking pairs: behavior of a neuronal model with STDP learning

    Neural Comput.

    (2008)
  • M. Tsodyks et al.

    Synchrony generation in recurrent networks with frequency-dependent synapses

    J. Neurosci.

    (2000)
  • X. Lin, P. De Wilde, Odor recognition with synchronization using integrate and fire neurons, in: Proceedings of...
  • Cited by (0)

    Xiaobin Lin is a Ph.D. candidate in the Department of Computer Science, Heriot-Watt University, Edinburgh, United Kingdom. He obtained a M.Sc. degree in Information and Intelligence from the Electronics and Electrical Engineering Department, University of Liverpool, United Kingdom, in 2005 and a B.S. degree in Computer Science from the Computer Science Department, Shenzhen University, Shenzhen, China, in 2004. His research interests include neural network topology and plasticity in the brain, communication between neurons and glial cells, system analysis and modeling.

    Philippe De Wilde is a professor in the Intelligent Systems Lab of the Department of Computer Science, School of Mathematical and Computer Sciences, Heriot-Watt University, Edinburgh, United Kingdom. He tries to discover biological and sociological principles that can improve the design of decision making and of networks. His research interests are: coupling of neuronal, glial, and cerebro-vascular networks; networked populations; coordination mechanisms for populations; group decision making under uncertainty; neural networks; neuro-economics; stability, scalability and evolution of multiagent systems. He is Research Fellow in British Telecom, 1994. Laureate, Royal Academy of Sciences, Letters and Fine Arts of Belgium, 1988. Senior Member of IEEE, Member of IEEE Computational Intelligence Society and Systems, Man and Cybernetics Society, and British Computer Society. He is an associate editor in IEEE Transactions on Systems, Man, and Cybernetics, Part B, Cybernetics.

    View full text