Function of attention in learning process in the olfactory bulb

Vol. 46 No.4SCIENCE IN CHINA (Series C)August 2003Function of attention in learning process in the olfactorybulbMA Baosheng (马宝生)', WANG Shunpeng (王顺鹏)', LIYan(李岩)',FENG Chunhua (冯春华)' & GUO Aike (郭爱克)121. Laboratory of Visual Information Processing, Institute of Biophysics, Chinese Academy of Sciences, Beiing 100101,China;2. Institute of Neuroscience, Chinese Academy of Sciences, Shanghai 200031, ChinaCorrespondence should be addressed to Guo Aike (email: akguo@ion.ac.cn)Received July 15, 2002Abstract It has been suggested that in the olfactory bulb, odor information is processed throughparallel channels and learning depends on the cognitive environment. The synapse s spike effec-tive time is defined as the effective time for a spike from pre-synapse to post-synapse, which varieswith the type of synapse. A learning model of the olfactory bulb was constructed for synapses withvarying spike effective times. The simulation results showed that such a model can realize themulti-channel processing of information in the bulb. Furthermore, the effect of the cognitive envi-ronment on the learning process was also studied. Different feedback frequencies were used toexpress different attention states. Considering the information' s multi-channel processing require-ment for learning, a learning rule considering both spike timing and average spike frequency isproposed. Simulation results showed that habituation and anti-habituation of an odor in the olfacto-ry bulb might be the result of learning guided by a common local learning rule but at different atten-tion states.Keywords: olfactory bulb, spike ffective time, learning rule, feedback, attention.DOI: 10.1360/0 1yc0301Biological experiments! 1- 8] have shown that in the olfactory bulb, odor information is encod-ed into spatio-temporal patterns and processed in multi- channels. In the bulb, the information isdivided into basic information and fine information, which are encoded into average spiking fre-quency (long- term pattern) and synchronization of spikes[1- 8] respectively. Compared with otherstructures in which information is encoded into spatio-temporal patterns, the olfactory bulb' sstructure is rather simple. Experiments have also shown that coding of an odor in the olfactorybulb was plastic and adjusted by the attention mechanism19,10]. Due to its simple structure and typi-cal learning and coding, the olfactory bulb model is suitable for the study of learning and memory.How is multi-channel information processing in the olfactory bulb carried out? Informationin the neural system is accomplished by the spikes 01中国煤化Iikefromapre-synapse to its post-synapse includes two attributes: inte.MYHC N M H Gthe criterions ofcoding and spike effective time are comparable, the spike effective time must be considered. Thispaper proposes a concept of spike effective time, defined as the effective time of a pre-synapticNo.4FUNCTION OF ATTENTION IN LEARNING PROCESS IN OLFACTORY BULB359spike to the post-synaptic neuron' s potential. Odor information is divided into molecular informa-tion (basic information) and the ratio information of the molecules (fine information). If these twokinds of information are processed by synapses with quite different spike effective time, it willbuild multiple channels for information processing.Multi channel coding requires that the learning rule must consider both the spike timing andthe average spike frequency. This paper purposes an antisymmetric learning rule for models basedon spike neurons: the modification of synapses relies on pre- and post- synaptic spike timing. Indetail, the synapse is modified only when the pre- and post- synapse spikes are within a learningwindow (5一-20 ms), the modification' s direction depends on the sequence of a pair of spikes, andthe modification' s size depends on the interval between the pre- and post- synaptic spikes. Thislearning rule has the following characteristics: 1. locality: a synapse' s modification only dependson its pre- and post- neuron spikes; 2. timing sensitivity: the timing of pre- and post- synapticspikes decides the size and direction of the modification; 3. interactivity: the synapse s modific a-tion is the result of interactivity between the pre- and post- neurons, i.e. only if both pre- and post-neurons spike can the synapse be modified. In this study, with a computer model of olfactory bulb,analysis proved that such a local learning rule can induce different global learning results at dif-ferent attention states. In other words, the habituation and anti-habituation of an odor are based onthe meaning of the odor to the creature. Here, the attention state reflects the effect of a cognitiveenvironment on the learning process. The attention state is brought about by feedback from thehigher cortex area, implying that the attention state might be carried out by the frequency of thefeedback spikes. Data from some related biological experimentsl11.121 suggested that the frequencyin the cognitive process at the attention state is higher than that in the cognitive process at the noattention state.This paper presents studies of the function of the synapse' s spike effective time in coding andlearning, as well as the role of attention in global learning.1 ModelBased on anatomical data, we constructed a simplified model of the olfactory bulb, whichwas composed of 16 periglomerular cells, 16 mitral cells and 160 granule cells (fig. 1). The mitralcells were arranged as a one-dimensional circle. Each mitral cell interacted with 10 surroundinggranule cells, and four mitral cells composed an olfactory glomerulus. The mitral cells within aglomerulus interacted with each other through inhibitory synapses, and the mitral cells from dif-ferent glomeruli interacted with each other through granule cells. The sensory neurons projectedinto mitral cells and periglomerular cells, and those projecting into a glomerulus had at least onecommon odor receptor. The feedback from the highe中国煤化Ie granule cells,which affected the bulb s output by adjusting the granulMYHCNM HGA neuron spikes when its membrane potential is over certain threshold, described as360SCIENCE IN CHINA (Series C)Vol.46ONOlfactory nerves→Glomerular layerPCfMMitral cell layerG↓Granule cell layerLOTH Olfactory cortex ]Fig. 1. Network structure of the model for the olfactory bulb. The network was composed of 3 layers: the periglomerularcell layer (PG), the mitral cell layer (M) and the granule cell layer (G). Each glomerulus was composed of 4 mitral cells,with the olfactory sensory neurons having the same odor receptor projected to mitral cells and periglomerular cells withina glomerulus, and mitral cells within a glomernulus interacting with each other. The mitral cells project to the granule cellswith excitatory synapses and receive inhibitory feedback from the granule cells. The feedback from the olfactory cortexprojects to the granule cells.口, Excitatory synapse; , inhibitory synapse.V.i =h(ux;-0;), x=m,p,8,(1)where xi is the x layer ith neuron' s threshold for spiking, h(x) is step function, ifx > 0, itis 1,otherwise 0.A spiking neuron' s membrane potential declines quickly. The ith x neuron' s potential de-clining speed when it is spiking is described asuU =B@Ee0-0)A,x=m,p, 8, .(2)where t is the time，k is the timing of the kth spike of the spike serial; Bm, B, and Bg are the am-plitudes of the declining speed during spiking; andis the time coefficient of a membrane poten-tial' s decline induced by spiking, reflecting the refractory period, set as 2 ms. Bm, B, and Bg wereset large enough to have the neurons in the model behave as spike neurons (Bm= B,= Bg = 2000ms l), so that the spiking neuron' s membrane potential declined below threshold very quickly (<1ms).The effect of a spike from pre-synapse to post-synapse is described asF5wi;() =wssjV.,(t)dt中国煤化工(3)CHCNMHGwhere Ty. i is the synapse' s spike effective time. When tne two neurons connecred by the synapseare within a glomerulus, it is 100 ms, otherwise 5 ms; W.v. j and dev. j are the weight and delay ofNo.4FUNCTION OF ATTENTION IN LEARNING PROCESS IN OLFACTORY BULB361the synapse from the jth y neuron to the ith x neuron.When the spike effective times of different synapses vary greatly (e.g. different magnitudes),they will decide the role of the synapse in coding. On a long-term scale, the input caused by spikesfrom the short-term-effective synapses will be eclipsed by the input caused by spikes from thelong-term-effective synapses. Thus the long-term output patterns mainly rely on spikes from long-term-effective synapses. On a short-term scale, because the input caused by spikes from short-term-effective synapses and long-term-effective synapses are comparable, the short-term outputpatterns are the result of both the spikes from long-term-effective synapses and short- term-effective synapses.Based on the olfactory bulb' s special structure, the multi-channel processing of informationcan be accomplished by setting the synapse' s spike effective time. As mentioned before, thoseolfactory sensory neurons projecting into a single glomerulus have the same odor receptor. If onlythe interaction within a glomerulus is considered, the output patterns rely on only the odor' s mo-lecule types; if the interaction between glomeruli is considered also, the output patterns rely on themixture ratio of the odor molecules also. In this model, only the synapses within a glomerulus arelong-term-effective, so the long-term output of mitral cells encodes only the information of theodor molecule types, while the mixture ratio of the molecules can only be ancoded by the short-term patterns (spike timing). Considering the element of the time scale, the model effectively rep-resented the multi-channel processing of odor information.One of the inputs to the bulb comes from olfactory sensory neurons. Each glomerulus re-ceived the exciting input from the sensory neurons that have the same odor receptor. Im. ;(t) andIp.;(t) are inputs from the olfactory sensory neurons to the ith mitral cell and pre glomerulus cell.For the odor X(X1, X2, X3, X4), the ratio of molecules are (n (X1)，(X2)， (X3),λ (X4)). Imi(t)and Ip,; (t) are defined asm,;()=Ip;()=x2 ,^(Xn );(X,),(4)where x is the density of the odor (x = 10); i is used to describe the speciality of the neurons toodor molecules. When the ith mitral cell is sensitive to the odor molecule X ; (X,) =1, otherwise0.The feedback from the olfactory cortex projects into the granule cells, and the feedback inputon the ith granule cells isIg,()=1.(sin20v() +1).Setting Wgm, i and I, we can make a granule cell spike only when the following conditions aresatisfied: () it receives its pre-synaptic mitral cell inpu中国煤化工(i) the feedbackreaches its peak. Such a granule cell is a kind of coincid:TYHCNMH Gncy of feedbackrelies on the attention state, which is set to 200 Hz at attention state, or 50 Hz at no attention state.Noticing the synapses from mitral cells to granule cells are all within a glomerulus, the spike ef-362SCIENCE IN CHINA (Series C)Vol.46fective time of these synapses is 100 ms. At attention state, the interval between the granule cellspike and its pre-synaptic cell spike is in the time window of 0- -5 ms; while at no attention state,the interval is mostly in the time window of 5- -20 ms[141.According to the bulb s structure, mitral cells, periglomerular cells and granule cells' statefunction are defined as:dum.idi=-m,n-umn +E E Fuwu()+1m.()+Em， x=m,p,g,(6)dup.ltA,upi-upr +1(0,(7)dug. ;=-Agu. ;-ufre +Fgm () +1。. ;(0,(8)dtwhere Um, ;(), up, ;() and ug,;() are membrane potentials, Am. A, and Ag are coefficients describingthe decline (Am=A,=Ag= 1 ms-), Em is the mitral cell' s self-enhanc ing efientlI5.16.The model used an anti-symmetric Hebb learning rule for spike neuronsl[171. The modificationof the weight of the synapse between spike neurons relies on spiking timing. The size is an inverseratio to the spikes interval and the direction of modification is decided by the sequence of thespikes, described asOw= εG(Ot),(9)where ε is the synapse' s modification coefficient, and△t is the difference between timing of post-and pre- synaptic cell spikes. G (O t) defines the relativity of the spikes:G(A)={|4rle (5,20),(10)[0 otherwise,where is 1 for an excitatory synapse and - -1 for an inhibitory synapse.The lower limit (5 ms) defined in the learning rule reflects the time needed for a pre-synapticspike to induce a post-synaptic cell to respond; the upper limit reflects the scale of informationrelativity. If the interval between the spikes of pre- and post- synaptic cells is less than 5 ms ormore than 20 ms, then the relation between these two spikes is not certain, this pair of spikeswould not be able to modify the synapse.For this paper, we used an odor composed of four kinds of molecules as input to study multi-channel encoding, the response to the same odor at the attention state (high frequency feedback)and at the no attention state (low frequency feedback), and learning.2 Simulation results中国煤化工MHCNMHGIn order to contrast with the biological experiments, we pruviueu 4 gioups of simulation re-sults: (i) responses induced by 2 similar odors in the bulb; (i) responses induced by an odor beforeNo.4FUNCTION OF ATTENTION IN LEARNING PROCESS IN OLFACTORY BULB363and after blocking interaction from granule cells to mitral cells; (ii) responses induced by an odorat high frequency feedback and low frequency feedback; and (iv) responses induced by an odorbefore and after learning process.The simulation results coincided well with the biological experiment rsult2-35.910, andproved: (i) in the bulb, synapses having different spike effective times can build multiple channelsfor odor information processing; (ii) the different responses induced by the same odor in the bulbmight be caused by the attention state; (ii) with a local learning rule, the bulb can use the attentionstate to produce different global learning results. In detail, first, the similar odors induce similarlong-term patterns, but short-term patterns (spike timing synchronization) are quite different. Sec-ond, the response induced by an odor at high frequency feedback is stronger than that at low fre-quency feedback. Third, the learning process at high frequency feedback strengthens the response,while the learning process at low frequency feedback weakens the response. Finally, after alearning process under high frequency feedback, synchronization is enhanced, while after alearning process under low frequency feedback, there is no obvious difference.At the scale of 100 ms, interactions through short-term (5 ms) spike effective synapses areeclipsed by interactions through long-term (100 ms) spike effective synapses (fig. 2). Because theolfactory sensory neurons projecting to the same glomerulus have the same odor receptor, andonly the synapses within a glomerulus are long-term spike effective, patterns induced by similarodors at this time scale (100 ms) are similar too, while the synapses from granule cells and mitralcells affect only the short-term pattern.Attention state affects the response s strength in the bulb. The responses induced by an odorat the attention state (high frequency feedback, 200 Hz) and at the no attention state (low frequen-cy feedback, 50 Hz) are shown in fig. 3. Both strength and range of responses at the attention stateare stronger than at the no attention state. Meanwhile, the attention state responses include all theresponses found in the no attention state, as shown in fig. 3. This was effectively simulated, andthe simulation results coincided well with biological experiment results.The biological experiments suggested that odor learning is adjusted by the attention statel9.10.The model was used to study how this was carried out in the bulb. The learning results werechecked at the attention state after the odor was learned at different attention states for 2 s.Considering computations, the attention state can affect learning results only by adjusting thesynapses from granule cells to mitral cells (fig. 4). The biological studies showed that a mitralcell' s input came from olfactory sensory neurons, other mitral cells and granule cells. The modifi-cation of these synapses from these three kinds of cells determined the output patterns after learn-ing. Because of the relativity of spikes, the synapses from olfactory sensory neurons to mitral cellswould be enhanced in any case. Because of the symm中国煤化Iti-symmetry oflearning rule, the learning of synapses between mitralYHCNMHGonofspikesbuthad nothing to do with the global strength of response. In fact, only the learning process of synap-ses between granule cells and mitral cells relied on the attention state directly. At the attention364SCIENCE IN CHINA (Series C)Vol.46state, the granule cells spiked immediately after a pre-synaptic mitral cell spike (<5 ms), so thestrength of the synapses did not change ( eq. (9)). Therefore, learning at the attention statestrengthened the response induced by the odor. At the no attention state, most of the spikes of a0tExperiment 101152Experiment2营shh202:Experiment 3(a)。105Time X 100/ms2.0 r.5-ExperimentI音1.0-1I|」|l山.050550 6507580859095100.0-Experiment20.5-业LII0.050 s5606507808590951001.5-Experiment3告1.0-(b)6065707580859095;10TimeX 5/msFig. 2. The long-term (100 ms) and short-term (5 ms) responses induced by two similar odors before and after thesynapses from granule cells to mitral cells were blocked. Experiment 1 and experiment 2 are the responses of odor Aand odor B before the synapses were blocked. Odor A and odor B中国煤化工olecle, but theratios were dfferent. Experiment 3 is the response induced by odorn. (a) Long-termpattemns (100 ms); (b) short-term pttrnrs (5 ms). The odor input beHCNMHGTheplotsshowthat the long-term patterns are similar but the short-term patterns are qutc unnerelnl. 1m1s 1ess unal the informationof the ratio of the odor molecules and the synapses from granule cells to mitral cells affected only the short termpatterns of mitral cells output, but was unrelated to long-term patterns.No.4FUNCTION OF ATTENTION IN LEARNING PROCESS IN OLFACTORY BULB36511113141:16a)HmMmMm12b)Fig. 3. The response of mitral cells induced by the same odo中国煤化工ntion respec-tively. (a) and (b) are the responses at the attention state and the n) represent theidentity numbers of mitral cells. Criterion: horizontal, 100 ms;:fYHCNMHGrthefedbackspike serial was 200 Hz at the attention state and 50 Hz at the no attention state (subsequent figures correspond tothis). As shown in the plot, the response at the attention state is stronger than that at no attention state.366SCIENCE IN CHINA (Series C)Vol.4601156a)_ImIn_m_mmImD112Fig. 4. Odor leamning in the bulb is adjusted by feedback. Thes中国煤化工k on the long-term response induced by odor A. After 2 s of learming, output pa:YHCNMHGattention statewere checked. (a) Response after attention-state learning; (b) respPing. Criterion:horizontal, 100 ms; vertical, 100 Hz. The plots show that the response after learning at high frequency feedback(attention state) was strengthened and the response after learning at low frequency feedback (no attention state)was weakened.No.4FUNCTION OF ATTENTION IN LEARNING PROCESS IN OLFACTORY BULB367granule cell and its pre-synaptic mitral cell fell into a learning window (5一20 ms), and the in-hibitory interaction from granule cells to mitral cells increased noticeably. When this inhibitoryinteraction increased faster than the excitatory interaction from the olfactory sensory neurons, thelearning process caused a response decline. The simulation results showed that learning at theattention state enhanced the response globally, while learning at the no attention state reduced theresponse strength and distribution. This coincided well with the previous biological exper i-mentsl9.10]. In other words, the simulation results proved that the odor' s habituation and in-habituation might be determined by feedback frequency.In the bulb, the synchronization of spikes plays an important role in information coding. Thefigures show synchronization patterns after learning at different attention states. As shown in fig. 5,the learning process at the attention state noticeably strengthened the spike synchronization, and itshowed some periodicity also. In contrast, the learning process at the no attention state made no4.0(a3.5|3.02.0言1.5|1.00.810830850870890 900TimeX 2/ms85.0(c)7兰2.0亮:2830 850 870890 90Time X 2/msTime x 2/msFig. 5. Effect caused by different feedback on learning. After 2 s of learning, the short-term response was check-ed at the atention state. In order to clearly show the synchronization state anlv 200me dotaochnwn in the pposThe vertical dimension is the sum of the spikes of all 16 mitral C中国煤化工; the time. (a)Output patterns before learming; (b) output patterns after attent:YHCNMHGermsafterno-attention-state learning. After learming at the attention state, the synchronization ot spikes strengthened measurably.SCIENCE IN CHINA (Series C)Vol.463 DiscussionSince neural systems are dynamic, information encoded might not be invariable. The learningprocess of coding might improve the efficiency of the neural system. Logically, learning enhancesthe response to meaningful information and reduces the response to meaningless information. Inthis paper, synapse spike effective time, as well as the olfactory bulb' s special structure was usedto achieve odor information multi-channel processing, and the results proved that the odor’ s ha-bituation and in-habituation can be the result of learning guided by a common local rule at differ-ent attention states.As the interaction from the higher cortex area to the lower cortex area, attention can only becarried out by feedback, and different attention states are expressed by different feedbacks. As oneof the important properties, the frequency of the feedback spike serial can efficiently identifyvariation in feedback. In addition, biological experiments suggested that in the olfactory bulb, thefrequencies of EEG induced by the odors varied with the familiarity degrees of the odorl11,121.Therefore, it is biologically plausible to suppose that attention' s variation expressed by the varia-tion of the feedback spike serial S frequency.One of the most important characteristics of this model is the synapse' s spike effective time.For models based on the average frequency coding, the synapse' s spike effective time is not relat-ed to coding because the output time scale is much larger than the synapse s spike effective time.But for models based on spike timing coding, which is comparable to the scale of coding, the syn-apse' s spike effective time must be considered. This study examined the function of spike effec-tive times with a model in which the synapses had different spike effective times, and proved thatmultiple channels for information processing can be constructed by such synapses.A Hebb learning rule based on spike neurons was also proposed. Compared with other Hebblearning rules based on average frequency coding neuron!8 21, it stressed the function of spiketim ing. The window limit to learning showed the time relativity of pre- and post- synaptic spikeson coding; at the same time, because of the anti-symmetry of the learming rule, learning strength-ened the excitatory synapses on the pathway of information transferring and weakened the others,and the learning of inhibitory synapses enhanced the synchronization of spikes.The anti-symmetric learning rule of the model shows the function of the information' s rela-tivity on learning. For a creature, the aim of local learning is to increase the information transfer.The learning rule adopted in this paper achieves this aim: if the pre- and post- synaptic spikestransfer information, they will strengthen the excitatory synapses and weaken the inhibitory syn-apses; otherwise, they will weaken the excitatory synapses and strengthen the inhibitory synapses.Finally, and most importantly, this work examined how global learning w as accomplished bylocal learning. Global learning is composed of local lea中国煤化Irf local learning.Global learning enhances the efficiency of meaningfulYHC N M H Geakens meanin-gless information transfer based on judgment of the information' s relevance. For a certain network,the use of its output can be judged only by its higher layer. In other words, global learning is theNo.4FUNCTION OF ATTENTION IN LEARNING PROCESS IN OLFACTORY BULB369result of cooperation between local learning and feedback. In this paper, the attention state adjust-ed a granule cell' s sensitivity to its pre-synaptic mitral cell spike, thereby adjusting the modific a-tion of the synapses from granule cells to mitral cells. Because the synapses from granule cells tomitral cells play an important role in odor coding in the bulbl22 23), feedback will eventually affectthe odor's coding in the bulb, and the attention state in the learning process will adjust globallearning.References1. Laurent, G., Dynamical representation of odors by oscillating and evolving neural assemblies, Trends in Neurosciences,1996, 19: 489- -496.2. Laurent, G., Wehr, M, Davidowitz, H., Temporal representations of odors in an olfactory network, J. Neuroscience, 1996,16: 3837- -3847.3. Laurent, G., Macleod, K, Stopfer, M. et al, Spatiotemporal structure of olfactory inputs to the mushroom bodies, Leam-ing & Memory, 1998, 5: 124- 132.4. Laurent, G, A systems perspective on early olfactory coding. Science, 1999, 286: 723- -728.5. MacLeod, K., Laurent, G., Distinct mechanisms for synchronization and temporal patterning of odor-encoding neural as-semblies, Science, 1996, 274: 976- 979.6. Mori, K., Relation of chemical structure to specificity of response in olfactory glomeruli, Current Opinion in Neurobiology, 1995, 5: 467- -474.7. Mori, K., Nago, H., Yohihara, Y., The olfactory bulb: Coding and processing of odor molecule information, Science, 1999,286.711- 715.8. Stopfer, M., Bhagavan, S., Smith, B. H. et al, Impaired odour discrimination on desynchronization of odour-ecoding neu-ral assemblies, Nature, 1997, 390: 70- 74.9. Buonviso, N., Chaput, M., Olfactory experience decreases responsiveness of the olfactory bulb in the adult rat, Neurosci-ence, 2000, 95(2): 325- 332.10. Faber, T, Joerges, J, Menzel, R., Associative learing modifies neural representations of odors in the insect brain, NatureNeuroscience, 1999, 2: 74- 78.| 1. Wilson, D. A., Sullivan, R. M.. Leon, M.. Odor familiarity alters mitral cell response in the olfactory bulb of neonatal rats,Brain Research, 1985, 22:314←-317.12. Gray, C. M., Skinner, J. E., Field potential response change in the rabbit olfactory bulb accompany behavioral habituationduring the repeated presentation of unreinforced odors, Experimental Brain Research, 1998, 73: 189- -197.13. Hiroshi Fuji, Hiroyuki Ito, Kazuyuki Aihara et al, Dynamical cell assembly hypothesis theoretical pssbilit of spatio-temporal coding in the cortex, Neural Networks, 1996, 9(8): 1303- -1350.14. Hendin, O., Horn, D.. Tsodyks, M V., Associative memory and segmentation in an osillatory neural model of the olfact 0-ry bulb, Journal of Computational Neuroscience, 1998, 5(2): 157- -169.15. Nicoll, R. A., Jahr, C. E, Self-excitation of olfactory bulb neurons, Nature, 1982. 296(5856): 441- 444.416. Davison, A. P., Feng, J, Brown, D.. Structure of lateral inhibition in an olfactory bulb model, Lecture Notes in ComputerScience, 999, 1606: 189- -196.17. Bi, G. Q, Poo, M. M.. Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synapticstrength, and postsynaptic cell type, J. Neuroscience, 1998, 18(24): 10464- -10472.8. Sejnowski, T. J., The book of Hebb, Neuron, 1999, 24(4): 773- 776.19. Tsien, J. Z, Linking Hebb' s coincidence-detection to memory formation, Current Opinion in Neurobiology, 2000, 10(2):266- 273.20. Turrigiano, G. G., Nelson, S. B., Hebb and homeostasis in neuronal plasticity, Current Opinion in Neurobiology, 2000,10(3): 358- 364.21. Viana, D.. Prisco, G.. Hebb synaptic platicit. Progess in Neurob中国煤化工22. Hendin, 0.. Horn, D.. Tsodyks, M. V.. The role of inhibition in an:YHC N M H Glfactory bulb, Joumalof Computational Neuroscience, 1997 4(2): 173- -182.23. Linster, C., Hasselmo, M.. Modulation of inhibition in a model of olfactory bulb reduces overlap in the neural representa-tion of olfactory stimuli, Behavioural Brain Research, 1997, 84(1-2): 117- -127.