Biol Cybern (2008) 98:459–478

DOI 10.1007/s00422-008-0233-1

REVIEW

Phenomenological models of synaptic plasticity based on spike timing

Abigail Morrison · Markus Diesmann ·
Wulfram Gerstner

Received: 16 January 2008 / Accepted: 9 April 2008

© The Author(s) 2008

Abstract Synaptic plasticity is considered to be the bio- 1 Introduction

logical substrate of learning and memory. In this document
we review phenomenological models of short-term and long- Synaptic changes are thought to be involved in learning,
term synaptic plasticity, in particular spike-timing dependent memory, and cortical plasticity, but the exact relation between
plasticity (STDP). The aim of the document is to provide a microscopic synaptic properties and macroscopic functional
framework for classifying and evaluating different models of consequences remains highly controversial. In experimental
plasticity. We focus on phenomenological synaptic models preparations, synaptic changes can be induced by specific sti-
that are compatible with integrate-and-fire type neuron mulation conditions defined through presynaptic firing rates
models where each neuron is described by a small number of (Bliss and Lomo 1973; Dudek and Bear 1992), postsynaptic
variables. This implies that synaptic update rules for short- membrane potential (Kelso et al. 1986; Artola et al. 1990),
term or long-term plasticity can only depend on spike timing calcium entry (Lisman 1989; Malenka et al. 1988), or spike
and, potentially, on membrane potential, as well as on the timing (Markram et al. 1997; Bi and Poo 2001).
value of the synaptic weight, or on low-pass filtered (tempo- Whereas detailed biophysical models are crucial to unders-
rally averaged) versions of the above variables. We examine tand the biological mechanisms underlying synaptic plasti-
the ability of the models to account for experimental data city, phenomenological models which describe the synaptic
and to fulfill expectations derived from theoretical considera- changes without reference to mechanism are generally more
tions. We further discuss their relations to teacher-based rules tractable and less computationally expensive. Consequently,
(supervised learning) and reward-based rules (reinforcement phenomenological models are of great use in analytical and
learning). All models discussed in this paper are suitable for simulation studies. In this manuscript, we will examine a
large-scale network simulations. number of phenomenological models with respect to their
compatibility with both experimental and theoretical results.
Keywords Spike-timing dependent plasticity · Short term In all cases, we consider a synapse from a presynaptic neuron
plasticity · Modeling · Simulation · Learning j to a postsynaptic neuron i. The strength of a connection
from j to i is characterized by a weight wi j that quantifies
A. Morrison · M. Diesmann the amplitude of the postsynaptic response, typically measu-
Computational Neuroscience Group,
RIKEN Brain Science Institute, Wako City, Japan
red as the height of the postsynaptic potential or the slope of
the postsynaptic current at onset. The conditions for synap-
M. Diesmann tic changes as well as their directions and magnitudes can
Bernstein Center for Computational Neuroscience, be formulated as ‘synaptic update rules’ or ‘learning rules’.
Albert-Ludwigs-University, Freiburg, Germany
Such rules can be developed from purely theoretical conside-
W. Gerstner (B) rations, or to account for macroscopic phenomena such as the
Laboratory of Computational Neuroscience, LCN, development of receptive fields, or based on findings from
Brain Mind Institute and School of Computer and Communication electrophysiological experiments manipulating firing rate or
Sciences, Ecole Polytechnique Fédérale de Lausanne (EPFL),
Station 15, 1015 Lausanne, Switzerland
voltage. In this manuscript, however, we restrict our scope to
e-mail: [email protected] rules which have been developed to account for the results

123
460 Biol Cybern (2008) 98:459–478

of experiments in which synaptic plasticity was observed as In this paper, we start with a review of some basic
a result of pre- and postsynaptic spikes (for more general experimental facts that could be relevant for modeling, follo-
reviews, see Dayan and Abbott 2001; Gerstner and Kistler wed by a list of theoretical concepts arising from fundamental
2002; Cooper et al. 2004). notions of learning and memory formation (Sect. 2). We then
For the classification of the synaptic plasticity rules, it review models of short-term plasticity in Sect. 3 and models
is important to specify the time necessary to induce such a of long-term potentiation/depression (LTP/LTD), in particu-
change as well as the time scale of persistence of the change. lar the spike-timing dependent form, in Sect. 4. Throughout
For both short-term and long-term plasticity, changes can the review we discuss spike-based plasticity rules from a
be induced in about 1 s or less. In short-term plasticity (see computational perspective, giving implementations that are
Sect. 3), a sequence of eight presynaptic spikes at 20 Hz appropriate for analytical and simulation approaches. In the
evokes successively smaller (depression) or successively lar- final sections we briefly mention reward driven learning rules
ger (facilitation) responses in the postsynaptic cell. The cha- for spiking neurons (Sect. 5) and provide an outlook toward
racteristic feature of short-term plasticity is that this change current challenges for modeling. The relevance of molecular
does not persist for more than a few hundred milliseconds: mechanisms and signaling chains (Lisman 1989; Malenka
the amplitude of the postsynaptic response recovers to close- et al. 1988) for models of synaptic plasticity (Lisman and
to-normal values within less than a second (Markram et al. Zhabotinsky 2001; Shouval et al. 2002; Rubin et al. 2005;
1998; Thomson et al. 1993). Badoual et al. 2006; Graupner and Brunel 2007; Zou and
In contrast to short-term plasticity, long-term potentiation Destexhe 2007), as well as the importance of the postsynaptic
and depression (LTP and LTD) refer to persistent changes of voltage (Kelso et al. 1986; Artola et al. 1990; Sjostrom et al.
synaptic responses (see Sect. 4). Note that the time neces- 2001), is acknowledged but not further explored.
sary for induction can still be relatively brief. For example,
in spike-timing-dependent plasticity (Bi and Poo 2001; Sjos-
trom et al. 2001), a change of the synapse can be induced by
2 Perspectives on plasticity
60 pairs of pre- and postsynaptic spikes with a repetition fre-
quency of 20 Hz; hence stimulation is over after 3 s. However,
Over the last 30 years, a large body of experimental results on
this change can persist for more than one hour. The final stabi-
synaptic plasticity has been accumulated. The most impor-
lization of, say, a potentiated synapse occurs only thereafter,
tant discoveries are summarized in Sect. 2.1. Simultaneously,
called the late phase of LTP (Frey and Morris 1997). An addi-
theoreticians have investigated the role of synaptic plasti-
tional aspect is that neurons in the brain must remain within
city in long-term memory, developmental learning and task-
a sustainable activity regime, despite the changes induced by
specific learning. The most important concepts arising from
LTP and LTD. This is achieved by homeostatic plasticity, an
this research are described in Sect. 2.2. Many of the plasti-
up- or down-regulation of all synapses converging onto the
city models employed in the theoretical approach were ins-
same postsynaptic neuron which occurs on the time scale of
pired by Hebb’s (1949) postulate that describes how synaptic
minutes to hours (Turrigiano and Nelson 2004).
connections should be modified:
The phenomenological models discussed in this manu-
script can be classified from a theoretical point of view as When an axon of cell A is near enough to excite cell B
unsupervised learning rules. There is no notion of a task to be or repeatedly or persistently takes part in firing it, some
solved, nor is there any notion of the change being ‘good’ or growth process or metabolic change takes place in one
‘bad’ for the survival of the animal; learning consists simply or both cells such that A’s efficiency, as one of the cells
of an adaptation of the synapse to the statistics of the activity firing B, is increased.
of pre- and postsynaptic neurons. This is to be contrasted with
reward-based learning, also called reinforcement learning In classical Hebbian models, this famous postulate is often
(SuttonandBarto1998).Inreward-basedlearningthedirection rephrased in the sense that modifications in the synaptic
and amount of change depends on the presence or absence transmission efficacy are driven by correlations in the firing
of a success signal, that may reflect the current reward or the activity of pre- and postsynaptic neurons. Even though the
difference between expected and received reward (Schultz idea of learning through correlations dates further back in the
et al. 1997). Reward-based learning rules are distinct from past (James 1890), correlation-based learning is now gene-
supervised learning since the success signal is considered rally called Hebbian learning. Most classic theoretical studies
as a global and unspecific feedback signal, that often comes represented the activity of pre- and postsynaptic neurons in
with a delay, whereas in supervised learning the feedback terms of rates, expressed as continuous functions. This has
is much more specific. In the theoretical literature, there led to a sound understanding of rate-based Hebbian lear-
exists a large variety of update rules that can be classified ning. However, rate-based Hebbian learning neglects the fine
as supervised, unsupervised or reward based learning rules. temporal structure between pre- and postsynaptic spikes.

123
Biol Cybern (2008) 98:459–478 461

Spike-based learning models for temporally structured (v) Plasticity depends on the postsynaptic potential (Kelso
input need to take this timing information into account (e.g. et al. 1986; Artola et al. 1990). If the postsynaptic neu-
Gerstner et al. 1993) which leads to models of spike-timing ron is clamped to a voltage slightly above rest during
dependent plasticity (STDP) (Gerstner et al. 1996; Kempter presynaptic spike arrival, the synapses are depressed,
et al. 1999; Roberts 1999; Abbott and Nelson 2000) that can while at higher depolarization the same stimulation
be seen as a spike-based generalization of Hebbian learning. leads to LTP (Artola et al. 1990; Ngezahayo et al.
The first experimental reports showing both long-term poten- 2000).
tiation and depression induced by causal and acausal spike (vi) On a slow time scale of hours, homeostatic changes of
timings on a time scale of 10 ms were published by Markram synapses may occur in form of rescaling of synaptic
and Sakmann (1995) and Markram et al. (1997), slightly response amplitudes (Turrigiano et al. 1994). These
after the theoretical work, however potentiation induced by changes can be useful to stabilize neuronal firing rates.
the pairing of EPSPs with postsynaptic depolarization on a (vii) Also on the time scale of hours, early phase LTP
time scale of 100 ms was demonstrated considerably earlier is consolidated into late phase LTP. During the
(Gustafsson et al. 1987). Timing in rate-based Hebbian lear- consolidation phase heterosynaptic interactions may
ning (although not spike-based) can be traced even further take place, probably as a result of synaptic tagging
back in the past (Levy and Steward 1983). From a conceptual and competition for scarce protein supply (Frey and
point of view, all spike-based and rate-based Hebbian lear- Morris 1997). Consolidation is thought to lead to long-
ning rules share the feature that only variables that are locally term stability of the synapses.
available at the synapse can be used to change the synaptic (viii) Distributions of synaptic strength (e.g., the EPSP
weight. These local elements that can be used to construct amplitudes) in data collected across several pairs of
such rules are listed in Sect. 2.3. neurons are reported to be unimodal (Sjostrom et al.
2001). At a first glance, this seems to be at odds with
experimental data suggesting that single synaptic
2.1 Experimental results
contacts are in fact binary (Petersen et al. 1998;
O’Connor et al. 2005).
The most important results from experiments on synaptic
(ix) Synapses do not form a homogeneous group, but dif-
plasticity with respect to the modeling of synaptic plasticity
ferent types of synapse have different plasticity pro-
are as follows:
perties (Abbott and Nelson 2000; Thomson and Lamy
2007). In fact, the same presynaptic neuron makes
(i) Short-term plasticity depends on the sequence of pre- connections to different types of target neurons with
synaptic spikes on a time scale of tens of milliseconds different plasticity properties for short-term (Markram
(Markram et al. 1998; Thomson et al. 1993). et al. 1998) and long-term plasticity (Lu et al. 2007).
(ii) Long-term plasticity is sensitive to the presynaptic
firing rate over a time scale of tens or hundreds of
Many other experimental features could be added to this list,
seconds. For example 900 presynaptic stimulation
e.g., the role of intracellular calcium, of NMDA receptors,
pulses at 1 Hz (i.e. 15 min of total stimulation time)
etc., but we will not do so; see Bliss and Collingridge (1993)
yield a persistent depression of the synapses, whereas
and Malenka and Nicoll (1993) for reviews. We emphasize
the same number of pulses at 50 Hz yields potentiation
that, given the heterogeneity of synapses between different
(Dudek and Bear 1992).
brain areas (plasticity has mainly been studied in visual or
(iii) Long-term plasticity depends on the exact timing of
somatosensory cortex and hippocampus) and between dif-
the pre- and postsynaptic spikes on the time scale of
ferent neuron and synapse types, we cannot expect that a
milliseconds (Markram et al. 1997; Bi and Poo 2001).
single theoretical model can account for all experimental
For example LTP is induced if a presynaptic spike
facts. In the next section, we will instead consider which theo-
precedes the postsynaptic one by 10 ms, whereas LTD
retical principles could guide our search for suitable plasticity
occurs if the order of spikes is reversed. In this context
rules.
it is important to realize that most experiments are
done with repetitions of 50–60 pairs of spikes whereas
a single pair has no effect. 2.2 Theoretical concepts
(iv) STDP depends on the repetition frequency of the pre-
post spike-pairings. In fact, 60 pairings pre-before- Synaptic plasticity is held to be the basis for long-term
post at low frequency have no effect, whereas the same memory, developmental learning, and task-specific learning.
number of pairs at a repetition frequency of 20 Hz gives From a theoretical point of view, synaptic learning rules
strong potentiation (Sjostrom et al. 2001). should therefore provide:

123
462 Biol Cybern (2008) 98:459–478

(i) sensitivity to correlations between pre- and postsynap-

tic neurons (Hebb 1949) in order to respond to corre-
lations in the input (Oja 1982). This is the essence of
all unsupervised learning rules
(ii) a mechanism for the development of input selectivity
such as receptive fields (Bienenstock et al. 1982; Miller
et al. 1989), in the presence of strong input features.
This is the essence of developmental learning
(iii) a high degree of stability (Fusi et al. 2005) in the
synaptic memories whilst remaining plastic (Grossberg
Fig. 1 Implementation of plasticity by local variables: each spike
1987). This is the essence of memory formation and contributes to a trace x(t). The update of the trace is either by a fixed
memory maintenance value (top) or to a fixed value (bottom)
(iv) the ability to take into account the quality of task perfor-
mance mediated by a global success signal (e.g. neuro-
modulators, Schultz et al. 1997). This is the consequence, spike-based plasticity models may be construc-
essence of reinforcement learning (Sutton and Barto ted from a combination of the following terms:
1998).
(i) spontaneous growth or decay (a non-Hebbian zero-
order term)—this could be a small effect that leads to
These items are not necessarily exclusive, and the relative slow ‘homeostatic’ scaling of weights in the absence
importance of a given aspect may vary from one subsystem to of any activity
the next; for example, synaptic memory maintenance might (ii) effects caused by postsynaptic spikes alone indepen-
be more important for a long-term memory system than for dent of presynaptic spike arrival (a non-Hebbian first-
primary sensory cortices. There is so far no rule which exhi- order term). This could be an additional realization
bits all of the above properties; moreover, theoretical models of homeostasis: if the postsynaptic neuron spikes at a
which reproduce some aspects of experimental findings are high rate over hours, all synapses are down-regulated
generally incompatible with other findings. For example, tra- (iii) effects caused by presynaptic spikes, independent of
ditional learning rules that have been proposed as an explana- postsynaptic variables (another non-Hebbian first-
tion of receptive field development (Bienenstock et al. 1982; order term). This is typically the case for short-term
Miller et al. 1989), exhibit a spontaneous separation of synap- synaptic plasticity
tic weights into two groups, even if the input shows no or only (iv) effects caused by presynaptic spikes in conjunction
weak correlations. This is difficult to reconcile with experi- with postsynaptic spikes (STDP) or in conjunction
mental results in visual cortex of young rats where a unimo- with postsynaptic depolarization (Hebbian terms)
dal distribution was found (Sjostrom et al. 2001). Moreover (v) all of the above effects may depend on the current value
model neurons that specialize early in development on one of the synaptic weight. For example, close to a maxi-
subset of features cannot readily re-adapt later on. On the mum weight synaptic changes could become smaller.
other hand, learning rules that do produce a unimodal distri-
bution of synaptic weights (van Rossum et al. 2000; Rubin We note that the changes induced by pre- or postsynaptic
et al. 2001; Gütig et al. 2003; Morrison et al. 2007) do not spikes need not necessarily immediately affect the synaptic
lead to long-term stability of synaptic changes, as the trajec- weight. Alternatively, they may lead to an update of an inter-
tories of individual synaptic weights perform random walks. nal hidden variable which evolves with some time constant
Hence it appears that long-term stability of memory requires τ . Hence, the hidden variable implements

 a low-pass filter.
a multimodal synapse distribution (Toyoizumi et al. 2007; For example, let us denote by δ t − t f a spike of a neu-
Billings and van Rossum 2008) or additional mechanisms to ron occurring at time t f . Then an internal variable x can be
stabilize the synaptic weights contributing to the retention of defined with dynamics:
a memory item.  
dx x 
=− + Aδ t −t f , (1)
dt τ f
2.3 Locally computable measures t

such that it is updated with each spike by an amount A and

In a typical spiking network model, a neuron is characteri- decays between spikes with a time constant τ (see Fig. 1,
zed by its voltage (subthreshold) and its firing times (spikes/ top). If the time constant is sufficiently long and A = 1/τ ,
superthreshold). Mesoscopic measures such as population the hidden variable gives an online estimate of the mean firing
activity are not accessible for the individual synapses. As a rate in the spike train. Other variations in the formulation of

123
Biol Cybern (2008) 98:459–478 463

such a ‘trace’ left by a spike are possible that do not scale 3.1 Markram–Tsodyks Model
linearly with the rate. First, instead of updating by the same
amount each time, we may induce saturation, One well-established phenomenological model for fast
synaptic dynamics was originally formulated for depression
dx x    only in Tsodyks and Markram (1997) and later extended to
=− + A (1 − x− ) δ t − t f . (2) facilitating dynamics in Markram et al. (1998). Here, we dis-
dt τ f t cuss the formulation of the model presented in Tsodyks et al.
(2000).
For A < 1 the amount of increase gets smaller as the variable If neuron i receives a synapse from neuron j (see Fig. 2b),
x before the update (denoted by x− ) approaches its maximal the synaptic current (or conductance) in neuron i is wi j yi j (t),
value of 1. Hence the variable x stays bounded in the range where wi j is the absolute strength and yi j (t) is a scaling factor
0 ≤ x ≤ 1. An extreme case of saturation is given by A = 1, that describes the momentary input to neuron i. Dropping the
in which case the reset is always to the value of 1, regardless indices for the rest of this discussion, y evolves according to:
of the value of x just before. In this case, the value of the trace  
dx z f
x depends only on the time since the most recent spike (see = − u + x− δ t − tj (3)
Fig. 1, bottom). We will see in the following sections that dt τrec
dy y  
the idea of traces left by pre- or postsynaptic spikes plays a = − + u + x− δ t − tj
f
(4)
fundamental role in algorithmic formulations of short-term dt τI
and long-term plasticity. For example, in the case of Hebbian dz y z
= − (5)
long-term potentiation, traces left by presynaptic spikes need dt τI τrec
to be combined with postsynaptic spikes, whereas short-term
where x,y and z are the fractions of synaptic resources in the
plasticity can be seen as induced by traces of presynaptic f
recovered, active, and inactive states respectively, t j gives the
spikes, independent of the state of the postsynaptic neuron.
timing of presynaptic spikes, τI is the decay constant of PSCs
In principle, voltage dependence could be treated in a simi-
and τrec is the recovery time from synaptic depression. These
lar fashion, see, e.g., Brader et al. (2007), but we will focus in
equations describe the use of synaptic resources by each pre-
the following on learning rules for short-term and long-term
synaptic spike—a fraction u + of the available resources x is
plasticity that use spike timing as the relevant variable for
used by each presynaptic spike. The variable u + therefore
inducing postsynaptic changes.
describes the effective use of the synaptic resources of the
synapses, which is analogous to the probability of release in
the model described in (Markram et al. 1998). The notation
3 Short-term plasticity x− in the update equations (3) and (4) is intended to remind
the reader that the value of x just before the update is used. In
Biological synapses have an inherent dynamics, which facilitating synapses, u + is not a fixed parameter, but derived
controls how the pattern of amplitudes of postsynaptic res- from a variable u which is increased with each presynaptic
ponses depends on the temporal pattern of the incoming spike and returns to baseline with a time constant τfac :
spike train. Notably, each successive spike can evoke a res-
du u  
ponse in the postsynaptic neuron that is smaller (depression) =− + U (1 − u − ) δ t − t j
f
(6)
or larger (facilitation) than the previous one. Its time scale dt τfac
ranges from 100 ms to about a second. Fast synaptic dyna-
mics is firmly established in biological literature (Markram where the parameter U determines the increase in u with
et al. 1998; Gupta et al. 2000), and well-accepted models each spike. We note that the update is equivalent to the satu-
exist for it (Abbott et al. 1997; Tsodyks et al. 1998). Neuro- rated trace (2). The notation u − indicates that the value of u
transmitter is released in quanta of fixed size, each evoking is taken just before the update caused by presynaptic spike
a contribution to the postsynaptic potential of fixed ampli- arrival. However, in (3) and (4) we use the value u + just
tude; this is known as the quantal synaptic potential (Kandel after the update of the variable u. If τfac → 0, facilitation
et al. 2000). The release of an individual quantum is known is not exhibited, and u + is identical to U after each spike,
to be stochastic, but the details of the mechanism underlying as is the case with depressing synapses between excitatory
this stochasticity remain unclear. However, the following two pyramidal neurons (Tsodyks and Markram 1997). The model
phenomenological models describe the average response and described by Eqs. 3–6 gives a very good fit to experimen-
are therefore entirely deterministic. Both models use the idea tal results: compare Fig. 2a and c. However, it should be
of a ‘trace’ left by presynaptic spikes (see previous section), noted that the values for the model parameters, including the
but in slightly different formulations. time constants, are quite heterogeneous, even within a single

123
464 Biol Cybern (2008) 98:459–478

A B C

Fig. 2 Short-term plasticity in experiment and simulation. a Experi- c Simulation results. The membrane potential of the postsynaptic neuron
mental results from rat cortex in slice. The average amplitude of the i is shown for repeated stimulations mediated by a synapse implemen-
evoked postsynaptic potential in neuron i varies with each successive ting the Markram–Tsodyks model. Experimental results adapted from
spike of the presynaptic neuron j. Top panel: depression; bottom panel: Markram et al. (1998), simulations performed with NEST (Gewaltig
facilitation. b Schematic representation of the neuron configuration. and Diesmann 2007)

neural circuit (Markram et al. 1998). The biophysical causes the system is reset to a new initial condition. Moreover, the
of the heterogeneity are still largely unclear. amount of synaptic resources is conserved - note that the
We note that not only the usage variable u, but also the right-hand sides of (3)–(5) add up to 0—thus z can be elimi-
variable y in (4) is essentially a ‘trace’ very similar to the one nated from the system. Let the state of the synapse be given
defined in the preceding section. To see this we eliminate the by:
variable x which is possible since the total amount of synaptic ⎛ ⎞
resources is fixed (x + y + z = 1). Hence (4) becomes: u
  s = ⎝x ⎠,
dy y f y
= − + u + (1 − y− − z − ) δ t − t j ,
dt τI
with the dynamics of u given by (6), that of y by (4)
 and that

which is a modification of the saturated trace in (2), the f
difference stemming from the fact that an additional ‘inac- of x given by dx/dt = (1 − x − y)/τrec − u + x − δ t − t j .
tive’ state has been introduced. Let us now suppose that Between two successive presynaptic firing times t  and t  the
the life-time of the ‘inactive’ state is short, i.e. τrec  τI . state of the synapse evolves linearly. At t  , the state of the
Then z decays rapidly back to zero and the above equation synapse without the effects of the new spike can be calculated
becomes the standard saturated trace. Since x = 1 − y and as:
dx/dt = −dy/dt, the available synaptic resources have the


  s t
dynamics: s t = Pt ,
1
dx 1−x  
=
f
− u + x− δ t − tj , (7) where t = t  − t  is the time difference of the two spikes
dt τI and (s(t  ), 1)T is a four-dimensional vector. The closed form
expression of the propagator matrix is:
which implies that the variable x is reduced at each presy-
⎛ − τt
⎞
naptic spike and, in the absence of spikes, approaches an
e fac 0 0 0
asymptotic value of unity with time constant τI . ⎜ ⎟
⎜ − t − t ⎟
The model defined in (3)–(6) can be solved using the tech- Pt =⎜
⎜ 0 e − τt
rec
τI (e τrec −e τI )
1−e − τt
rec
⎟.
⎟
nique of exact integration (Rotter and Diesmann 1999) by ⎝ τI −τrec ⎠
exploiting the following observations. The system of diffe- − t
τI
0 0 e 0
rential equations (3)–(6) is essentially linear, because all pro-
ducts of state variables are multiplied with delta functions. The state of the synapse at t  is the sum of the linear evolution
Therefore, between each presynaptic spike the system can since t  and the non-linear modification of the state due to
be integrated linearly, and on the occurrence of each spike the new spike:

123
Biol Cybern (2008) 98:459–478 465



s t  = s t  + s0 , In the case of depression, activity at the synapse causes
Prel to be decreased by f D Prel :
where the initial conditions are given by:
d Prel P0 − Prel  
⎛ ⎞ =
f
− f D Prel δ t − t j , (9)
U (1 − u) dt τP
s0 = ⎝ −x [u + U (1 − u)] ⎠ .
x [u + U (1 − u)] where f D controls the degree of depression (with 0 ≤ f D ≤
1), and the factor Prel prevents the release probability from
Note that the updated value of u is used to update the variables becoming less than 0. Note that with an equilibrium value
x and y. This reflects the assumption that the effectivity of P0 = 1 (which is always possible) this equation is equiva-
resource use is determined not just by the history of the lent to the that of the simplified Tsodyks model without the
synapse but also by the arrival of the new presynaptic spike, inactive state (7).
thus ensuring a non-zero response to the first spike (Tsodyks
et al. 1998).
In many simulation systems synapse models are constrai- 4 Long-term plasticity (STDP)
ned to transmit a synaptic weight rather than a continuous
synaptic current. In such cases, the synaptic weight transmit- Experimentally reported STDP curves vary qualitatively
ted to the postsynaptic neuron is wi j y0 , assuming the post- depending on the system and the neuron type—see Abbott
synaptic neuron reproduces the dynamics of the y variable. and Nelson (2000) and Bi and Poo (2001) for reviews. It is
It is not necessary for the neuron to reproduce the y dyna- therefore obvious that we cannot expect that a single STDP
mics for each individual synapse; due to the linearity of y rule, be it defined in the framework of temporal traces outli-
between increments, all synapses with the same τI can be ned above or in a more biophysical framework, would hold
lumped together. This is the implementation used in NEST for all experimental preparations and across all neuron and
(Gewaltig and Diesmann 2007). If the postsynaptic neuron synapse types. The first spike-timing experiments were
also implements an exact integration scheme (for a worked perform by Markram and Sakmann on layer 5 pyramidal neu-
example see Morrison et al. 2007), the dynamics of y can rons in neocortex (Markram et al. 1997). In the neocortex, the
be incorporated into the propagator of the dynamics of the width of the negative window seems to vary depending on
postsynaptic neuron. layer, and inhibitory neurons seem to have a more symmetric
STDP curve. The standard STDP curve that has become an
3.2 Abbott model icon of theoretical research on STDP (Fig. 1 in Bi and Poo
1998) was originally found for pyramidal neurons in rat hip-
A simpler model was developed by Abbott et al. (1997), for pocampal cell culture. Inverted STDP curves have also been
a complete description see Dayan and Abbott (2001). In this reported, for example in the ELL system in electric fish. This
model, synaptic conductance is expressed as gs = g s Ps Prel , gives rise to different functional properties (Bell et al. 1997).
where g s is the maximum conductance, Ps is the fraction of
open postsynaptic channels and Prel is the fraction of presy- 4.1 Pair-based STDP rules
naptic sites releasing transmitter. Ps generates the shape of
the postsynaptic conductance, and will not be further consi- Most models of STDP interpret the biological evidence in
dered here. Facilitation and depression can both be mode- terms of a pair-based update rule, i.e. the change in weight of
led as presynaptic processes that modify Prel . In both cases, a synapse depends on the temporal difference between pairs
between presynaptic action potentials Prel decays exponen- of pre- and postsynaptic spikes:
tially with a time constant τP back to its ‘resting’ level P0 . In
w + = F+ (w) · exp (− |t| /τ+ ) if t > 0
the case of facilitation, a presynaptic spike causes Prel to be (10)
−
increased by f F (1 − Prel ): w = −F− (w) · exp (− |t| /τ− ) if t ≤ 0,
f f
  where t = ti − t j is the temporal difference between
d Prel P0 − Prel
=
f
+ f F (1 − Prel ) δ t − t j , (8) the post- and the presynaptic spikes, and F± (w) describes
dt τP the dependence of the update on the current weight of the
synapse. A pair-based model is fully specified by defining:
f
where t j is the timing of the presynaptic spikes, f F controls (i) the form of F± (w); (ii) which pairs are taken into consi-
the degree of facilitation (with 0 ≤ f F ≤ 1), and the factor deration to perform an update. In order to incorporate STDP
(1− Prel ) prevents the release probability from growing larger into a neuronal network simulation, it is also necessary to
than 1. Note that (8) is just a modification of the saturated specify how the synaptic delay is partitioned into axonal and
trace in (2) due to a nonzero ‘resting’ level. dendritic contributions.

123
466 Biol Cybern (2008) 98:459–478

A 100 B 300
100
75

PSC change [pA]

PSC change [%]
30
50 10
3
25 0
−3
0 −10
−30
−25 −100
−300
−50
30 100 300 1000 3000 30 100 300 1000 3000
Initial PSC [pA] Initial PSC [pA]

Fig. 4 a Percentage change in peak synaptic amplitude after perfor-

ming the Bi and Poo (1998) spike pairing protocol as a function of
initial synaptic amplitude. Data from Bi and Poo (1998). b As in a, but
the absolute weight change is plotted rather than the percentage weight
Fig. 3 Implementation of pair-based plasticity by local variables. The change, and a double logarithmic representation is used. Fitted additive
spikes of presynaptic neuron j leave a trace x j (t) and the spikes of the potentiation update, gray dashed curve; fitted multiplicative potentia-
postsynaptic neuron i leave a trace yi (t). The update of the weight wi j tion update, solid gray curve; fitted power law update (slopes 0.4 and
at the moment of a postsynaptic spike is proportional to the momen- −1), black curves. Figure adapted from Morrison et al. (2007)
tary value of the trace x j (t) (filled circles). This gives the amount of
potentiation due to pre-before-post pairings. Analogously, the update
of wi j on the occurrence of a presynaptic spike is proportional to the proportional to the trace x j left by previous presynaptic
momentary value of the trace yi (t) (unfilled circles), which gives the spikes:
amount of depression due to post-before-pre pairings  
  f
wi+j ti = F+ wi j x j ti
f
(13)
 
  
wi−j t j = −F− wi j yi t j ,
f f
A pair-based update rule can be easily implemented with (14)
two local variables: one for a low-pass filtered version of the or alternatively:
presynaptic spike train and one for the postsynaptic spike
dwi j
  
train. The concept is illustrated in Fig. 3. Let us consider the = −F− wi j yi (t) δ t − t j
f
synapse between neuron j and neuron i. Suppose that each dt  

 f
spike from presynaptic neuron j contributes to a trace x j at +F+ wi j x j (t) δ t − ti . (15)
the synapse:
A pseudo-code algorithm along these lines for simulating
xj    arbitrary pair-based STDP update rules that is suitable for
dx j f
=− + δ t − tj , (11) distributed computing is given in Morrison et al. (2007).
dt τx f
tj Depending on the definition of the trace dynamics (accu-
mulating or saturating, see Sect. 2.3), different spike pairing
schemes can be realized. Before we turn to the consequences
f of these subtle differences (Sect. 4.1.2) and the implemen-
where t j denotes the firing times of the presynaptic neuron.
In other words, the variable is increased by an amount of one tation of synaptic delays (Sect. 4.1.3), we now discuss the
at the moment of a presynaptic spike and decreases exponen- choice of the factors F+ (w) and F− (w), i.e. the weight
tially with time constant τx afterwards; see the discussion of dependence of STDP.
traces in Sect. 2.3. Similarly, each spike from postsynaptic
neuron i contributes to a trace yi : 4.1.1 Weight dependence of STDP

dyi yi   f
 The clearest experimental evidence for the weight depen-
=− + δ t − ti , (12) dence of STDP can be found in Bi and Poo (1998), see
dt τy f
ti Fig. 4a. Unfortunately, it is difficult to interpret this figure
accurately, as the unit of the ordinate is percentage change,
and thus not independent of the value on the abscissa. An
f
where ti denotes the firing times of the postsynaptic neu- additional confounding factor is that the timing interval used
ron. On the occurrence of a presynaptic spike, a decrease of in the spike pairing protocol varies considerably across the
the weight is induced proportional to the momentary value data. However, even given these drawbacks, the rather flat
of the postsynaptic trace yi . Likewise, on the occurrence of dependence of the percentage weight change for depression
a postsynaptic spike a potentiation of the weight is induced (w/w ≈ constant) suggests a multiplicative dependence

123
Biol Cybern (2008) 98:459–478 467

A B C

Fig. 5 a Histogram showing the equilibrium distribution of synaptic times and mean input rate rin = 10 Hz. These cells converge on a single
efficacies using the additive STDP rule with τ = 10 ms, α = 1.05, conductance-based integrate and fire cell with parameters τm = 20 ms,
λ = 0.005. The upper two histograms show the behavior of a single τs = 5 ms, Vs = 5, and gs = 0.01. b As in a, but for multiplicative
synapse. In the top panel the presynaptic neuron is repeatedly stimulated STDP. Adapted from Rubin et al. (2001). c Equilibrium weight distri-
before the postsynaptic neuron. In the middle panel the timing relation bution in logarithmic gray scale as a function of the weight dependence
is reversed. The bottom histogram is the distribution of synaptic effica- exponent µ for an integrate and fire neuron driven by 1,000 uncorrelated
cies in a network of N = 1,000 excitatory cells with Poisson activation Poisson processes at 10 Hz. Adapted from Gütig et al. (2003)

of depression on the initial synaptic strength (w ∝ w). For F− (w) = λαw µ , F+ (w) = λ(1 − w)µ . A choice of µ = 0
potentiation the picture is less clear. results in additive STDP, a choice of µ = 1 leads to multi-
Instead of plotting the percentage weight change, Fig. 4b plicative STDP, and intermediate values result in rules which
shows the absolute weight change in double logarithmic have an intermediate dependence on the synaptic strength.
representation. The exponent of the weight dependence can Gütig et al. (2003) further demonstrated that the unimodal
now be determined from the slope of a linear fit to the data, distribution is the rule rather than the exception for update
see Morrison et al. (2007) for more details. A multiplicative rules of this form. A bimodal distribution is only produced
update rule (F− (w) ∝ w) is the best fit to the depression by rules with a very weak weight dependence (i.e. µ  1).
data but a poor fit to the potentiation data. The best fit to the Moreover, the critical value for µ at which bimodal distri-
potentiation data is a power law update (F+ (w) ∝ w µ ). The butions appear decreases as the the effective population size
quality of an additive update (F+ (w) = A+ ) fit is between Nr τ increases, where N is the number of synapses conver-
the power law fit and the multiplicative fit. ging onto the postsynaptic neuron, r is the rate of the input
spike trains in Hz and τ is the time constant of the STDP win-
4.1.1.1 Unimodal versus bimodal distributions dow (assumed to be equal for potentiation and depression).
The choice of update rule can have a large influence on the Figure 5c shows the equilibrium distributions as a function
equilibrium weight distribution in the case of uncorrelated of µ for N = 1,000, r = 10 Hz and τ = 0.02 s. µcrit is
Poissonian inputs. This was first demonstrated by Rubin et al. already very low for this effective population size. Because
(2001), see Fig. 5. Here, the behavior of an additive STDP of the high connectivity of the cortex, we may expect that the
rule (F+ (w) = λ, F− (w) = λα, where λ  1 is the learning effective population size in vivo would be an order of magni-
rate and α an asymmetry parameter) is compared with the tude greater, and so the region of bimodal stability would
behavior of a multiplicative STDP rule (F+ (w) = λ(1 − w), be vanishingly small according to this analysis. It is worth
F− (w) = λαw, with w in the range [0, 1). In the lowest his- noting that in the case that a sub-group of inputs is correlated,
tograms, the equilibrium distributions are shown for a neu- a bimodal distribution develops for all values of µ, whereby
ron receiving 1,000 uncorrelated Poissonian spike trains at the synaptic weights of the correlated group become stronger
10 Hz. In the case of additive STDP, a bimodal distribution than those of the uncorrelated group (data not shown—see
develops, whereas in the case of multiplicative STDP, the Gütig et al. 2003). In contrast to a purely additive rule, the
equilibrium distribution is unimodal. Experimental evidence peaks of the distributions are not at the extrema of the permit-
currently suggests that a unimodal distribution of synaptic ted weight range. Moreover, the bimodal distribution does not
strengths is more realistic than the extreme bimodal distri- persist if the correlations in the input are removed after lear-
bution depicted in Fig. 5a, see, for example, Turrigiano et al. ning. A unimodal distribution for uncorrelated Poissonian
(1998) and Song et al. (2005). Gütig et al. (2003) exten- inputs and an ability to develop multimodal distributions in
ded this analysis by regarding additive and multiplicative the presence of correlation is also exhibited by the additive/
STDP as the two extrema of a continuous spectrum of rules: multiplicative update rule proposed by van Rossum et al.

123
468 Biol Cybern (2008) 98:459–478

(2000): F+ (w) = λ, F− (w) = λαw; and by the power

law update rule proposed by Morrison et al. (2007) and also
Standage et al. (2007): F+ (w) ∝ λw µ , F− (w) ∝ λαw.

4.1.1.2 Fixed point analysis of STDP update rules

An insight into the similarity of behavior of all of these for-
mulations of STDP with the exception of the additive update
rule can be obtained by considering their fixed point struc-
ture. Equation (10) gives the updates of an individual synap-
tic weight. If the pre- and postsynaptic spike trains are sto-
chastic, the weight updates can be described as a random
walk. Using Fokker–Planck mean field theory, the average
rate of change of synaptic strength corresponds to the drift
of the random walk, which can be expressed in terms of the
correlation between the pre- and postsynaptic spike trains
(Kempter et al. 1999; Kistler and van Hemmen 2000; Kemp-
ter et al. 2001; Rubin et al. 2001; Gütig et al. 2003). Writing
  f
the presynaptic spike train as ρ j = t f δ t − t j and the Fig. 6 The fixed point structure of STDP: The synaptic drift ẇ scaled
j 
 f by the square of the pre- and postsynaptic firing rate ν is plotted as a
postsynaptic spike train as ρi = t f δ t − ti , the mean function of the synaptic weight w for various formulations of STDP.
  i
Model parameters chosen for visual clarity as only the qualitative beha-
rates are νi/j = ρi/j . Assuming stationarity, the raw cross-
vior is relevant. Additive, gray dash-dotted curve; multiplicative, gray
correlation function is given by dashed curve; van Rossum, gray curve; intermediate Gütig, black curve;
power law, dashed black curve
 
 ji (t) = ρ j (t)ρi (t + t) t (16)

i.e. averaging over t while keeping the delay t between ẇ

= −F− (w)τ− + F+ (w) τ+ . (18)
the two spike trains fixed. The synaptic drift is obtained by ν2
integrating the synaptic weight changes given by (10) over
t weighted by the probability, as expressed by (16), of the The fixed points of the synaptic dynamics are given by defi-
temporal difference t occurring between a pre- and post- nition by ẇ = 0, and therefore also by ẇ/ν 2 = 0. Figure 6
synaptic spike: plots (18) for a range of w and a variety of STDP models. In
0 all cases except for additive STDP the curves pass through
ẇ = −F− (w) dt K − (t)  ji (t) ẇ/ν 2 = 0 at an intermediate value of w and with a nega-
tive slope, i.e. for weights below the fixed point there is a
−∞
net potentiating effect, and for weights above the fixed point
∞
there is a net depressing effect, resulting in a stable fixed
+F+ (w) dt K + (t)  ji (t) , (17)
point which is not at an extremum of the weight range. In the
0 case of additive STDP there is no such fixed point, stable or
where K ± (t) = exp(− |t| /τ± ), the window function of otherwise.
STDP. The behavior of the additive model can be assessed more
As we are only interested in the qualitative structure of accurately by relaxing the assumption that pre- and postsy-
fixed points rather than their exact location, we will simplify naptic spike trains can be described by independent Poisson
the analysis by assuming that the pre- and postsynaptic spike processes. Instead, we consider a very simple neuron model
trains are independent Poisson processes with the same rate, in which the output spike train is generated by an inhomo-
 
i.e. ρi = ρ j = ν and  ji (t) = ν 2 . We can therefore geneous Poisson process with rate νi (u i ) = [α u i − ν0 ]+
write: with scaling factor α, threshold
 ν0 and membrane poten-
 f
tial u i (t) = j wi j t − t j , where (t) denotes the time
ẇ = −F− (w)τ− ν 2 + F+ (w) τ+ ν 2 .
course of an excitatory postsynaptic potential generated by a
In general, the rate ν of a neuron is dependent on the weight presynaptic spike arrival. The notation [x]+ denotes a piece-
of its incoming synapses and so the right side of this equation wise linear function: [x]+ = x for x > 0 and zero otherwise.
cannot be easily determined. However, we can reformulate In the following we assume that the argument of our piece-
the equation as: wise linear function is positive so that we can suppress the

123
Biol Cybern (2008) 98:459–478 469

square brackets. Assuming once again that all input spike with a fixed point given by:
trains are Poisson processes with rate ν, the expected firing
rate of the postsynaptic neuron is simply: Css ν0
νFP = (20)
 N Cν ¯ − Css
νi = −ν0 + αν ¯ wi j ,
j and a time constant τν = (αν [N Cν ¯ − Css ])−1 . Note that
 stabilization at a positive rate requires that ν0 > 0 and C > 0.
where ¯ = (s)ds, the total area under an excitatory post- The first condition states that, in the absence of any input, the
synaptic potential. The conditional rate of firing given an neuron does not show any spontaneous activity, and this is
f
input spike at time t j is given by trivially true for all standard neuron models, including the
   integrate-and-fire model. The latter condition is equivalent
f
νi (t) = −ν0 + αν ¯ wi j + αwi j t − tj , to the requirement
 that the integral over the STDP curve be
j negative: ds[F+ (w)K + (s) − F− (w)K − (s)] = −C < 0.
Exact conditions for stabilization of output rates are given
 Since for constant input rates ν we
thus the postsynaptic spike train is correlated with the presy- in Kempter et al. (2001).
naptic spike trains. This term shows up as additional spike- have νi = ν0 + αν ¯ j wi j , stabilization of the output rate
spike correlations in the correlation function  ji . Hence, in implies normalization of the summed weights. Hence STDP
addition to the terms in (18), thesynaptic dynamics contains can lead to a control of total presynaptic input and of the
a term of the form ανw F+ (w) K + (s) (s)ds that is linear postsynaptic firing rate – a feature that is usually associated
rather than quadratic in the presynaptic firing rate (Kempter with homeostatic processes rather than Hebbian learning per
et al. 1999, 2001). With this additional term, (18) becomes se (Kempter et al. 1999, 2001; Song et al. 2000).
Note that the existence of a fixed point and its stability
ẇ
= νi [−F− (w)τ− + F+ (w) τ+ ] does not crucially depend on the presence of soft or hard
ν  bounds on the weight. Equations (18) and (19) can equate to
+αwF+ (w) K + (s) (s)ds. (19) zero for hard-bounded or or unbounded rules.

For the multiplicative models the argument hardly changes, 4.1.1.3 Consequences for network stability
but for the additive model it does. For C = F− (w)τ− − Results on the consequences of STDP in large-scale net-
F+ (w) τ+ > 0, the additive model has a fixed point which works are few and far between, and tend to contradict each
we find by setting the right-hand side of (19) to zero,i.e. other. Part of the reason for the lack of simulation papers
on this important subject is the fact that simulating such
0 = −Cνi + αw Css , networks consumes huge amounts of memory, is computa-
 tionally expensive, and potentially requires extremely long
where Css = F+ (w) K + (s) (s)ds denotes the contribu- simulation times to overcome transients in the weight dyna-
tion of the spike–spike correlations. In contrast to the curves mics which can be of the order of hundreds of seconds of
in Fig. 6, the slope at the zero-crossing is now positive, indi- biological time. A lack of theoretical papers on the subject
cating instability of the fixed point. This instability leads to can be explained by the complexity of the interactions bet-
the formation of a bimodal weight distribution that is typi- ween the activity dynamics of the network and the weight
cal for the additive model. Despite the instability of indivi- dynamics, although some progress is being made in this area
dual weights (which move to their upper or lower bounds), (Burkitt et al. 2007).
the mean firing rate of the neuron is stabilized (Kempter It was recently shown that power law STDP is compa-
et al. 2001). To see this we consider the evolution of the tible with balanced random networks in the asynchronous-

output rate dνi /dt = αν ¯  j dwi j /dt. Since dwi j /dt = irregular regime (Morrison et al. 2007), resulting in a
−Cνi ν + ανwi j Css and αν ¯ j wi j = νi + ν0 , we can unimodal distribution of weights and no self-organization
write: of structure. This result was verified for Gütig et al. (2003)
STDP for an intermediate value of the exponent (µ = 0.4).
dνi
= −αν 2 ¯ N Cνi + αν (νi + ν0 ) Css , Although it has not yet been possible to perform systematic
dt
tests, it seems likely that all the formulations of STDP with
where N is the number of synapses converging on the post- the fixed point structure discussed in Sect. 4.1.1.1 would give
synaptic neuron. Thus we have a dynamics of the form: qualitatively similar behavior. The results for additive STDP
seem to be more contradictory. Izhikevich et al. (2004) repor-
dνi (νi − νFP ) ted self-organization of neuronal groups, whereas the chief
=−
dt τν feature of the networks investigated by Iglesias et al. (2005)

123
470 Biol Cybern (2008) 98:459–478

seems to be extensive withering of the synaptic connections. A

In the former case, it is the existence of many strong synapses
which defines the network, in the latter, the presence of many
weak ones. This discrepancy may be attributable to different
choices for the effective stabilized firing rates (20) in com-
bination with different choices of delays in the network, see B
Sect. 4.1.3.

4.1.2 Spike pairing scheme

There are many possible ways to pair pre- and postsynaptic C

spikes to generate a weight update in an STDP model. In an
all-to-all scheme, each presynaptic spike is paired with all
previous postsynaptic spikes to effect depression, and each
postsynaptic spike is paired with all previous presynaptic
spikes to effect potentiation. This is the interpretation used Fig. 7 Examples of nearest neighbor spike pairing schemes for a pre-
for the fixed point analysis in Sect. 4.1.1.1 and can be imple- synaptic neuron j and a postsynaptic neuron i. In each case, the dark
gray indicate which pairings contribute toward depression of a synapse,
mented using local variables as demonstrated in Sect. 4.1. In
and light gray indicate which pairings contribute toward potentiation.
a nearest neighbor scheme, only the closest interactions are a Symmetric interpretation: each presynaptic spike is paired with the
considered. However, there are multiple possible interpreta- last postsynaptic spike, and each postsynaptic spike is paired with the
tions of nearest neighbor, as can be seen in Fig. 7. Nearest last presynaptic spike (Morrison et al. 2007). b Presynaptic centered
interpretation: each presynaptic spike is paired with the last postsynap-
neighbor schemes can also be realized in terms of appropria-
tic spike and the next postsynaptic spike (Izhikevich and Desai 2003;
tely chosen local variables. The symmetric nearest-neighbor Burkitt et al. 2004: Model II). c Reduced symmetric interpretation: as
scheme shown in Fig. 7a can be implemented by pre- and in c but only for immediate pairings (Burkitt et al. 2004: Model IV, also
postsynaptic traces that reset to 1, rather than incrementing implemented in hardware by Schemmel et al. 2006)
by 1 as is the case for the all-to-all scheme. In the case of the
presynaptic centered interpretation depicted in Fig. 7B, the
data was best fit by a nearest neighbor interaction similar to
postsynaptic trace resets to 1 as in the previous example, but
Fig. 7c but giving precedence to LTP, i.e. a postsynaptic spike
the presynaptic trace must be implemented with a slightly
can only contribute to a post-before-pre pairing if it has not
more complicated dynamics:
already contributed to a pre-before-post pairing. However,
dx j xj        this result may also be due to the limitations of pair-based
1 − x− −
f f
=− + δ t − t − x δ t − ti , STDP models to explain the experimentally observed fre-
dt τx f
j j
f
j
tj ti quency dependence, see Sect. 4.2. More recently, Froemke
et al. (2006) demonstrated that the amount of LTD was not
f f
where t j and ti denote the firing times of the pre- and post- dependent on the number of presynaptic spikes following a
synaptic neurons respectively, and x − j gives the value of x j
postsynaptic spike, suggesting nearest-neighbor interactions
just before the update. In other words, the trace is reset to for depression as in Fig. 7c. However, the amount of LTP was
1 on the occurrence of a presynaptic spike and reset to 0 on negatively correlated with the number of presynaptic spikes
the occurrence of a postsynaptic spike. Similarly, the reduced preceding a postsynaptic spike. This suggests that multiple
symmetric interpretation shown in Fig. 7c can be implemen- spike pairings contribute to LTP, but not in the linear fashion
ted by pre- and postsynaptic ‘doubly resetting’ traces of this of the all-to-all scheme, which would predict a positive corre-
form. lation between the number of spikes and the amount of LTP.
It is sometimes assumed that the scheme used makes no Again, these results are good evidence for the limitations of
difference, as the ISI of cortical network models is typi- pair-based STDP rules.
cally an order of magnitude larger than the time constant
of the STDP window. However, this is not generally true 4.1.3 Synaptic delays
(Kempter et al. 2001; Izhikevich and Desai 2003; Morrison
et al. 2007). For a review of a wide variety of schemes and Up until now we have referred to t as the temporal diffe-
f
their consequences, particularly with respect to selectivity rence between a pre- and a postsynaptic spike, i.e. t = ti −
f
of higher-frequency inputs, see Burkitt et al. (2004). Expe- t j . However, many classic STDP experiments are expres-
rimental results on this issue suggest limited interaction bet- sed in terms of the temporal difference between the start of
ween pairs of spikes. Sjostrom et al. (2001) found that their the EPSP and the postsynaptic spike (Markram et al. 1997;

123
Biol Cybern (2008) 98:459–478 471

A B C

Fig. 8 Different partitions of synaptic delays and the resulting shift of delay is the same as the backpropagation delay. The synaptic raw cross-
the raw cross-correlation function as perceived at the synapse (black correlation function is identical to the somatic raw cross-correlation
curves) with respect to the raw cross-correlation function as percei- function. c All synaptic delay is dendritic, axonal delay is 0. The synap-
ved at the soma (gray curves). The cross-correlation functions shown tic raw cross-correlation function is shifted to the right by d. See Senn
are purely illustrative and do not result from a specific network model. et al. (2002) for a more detailed treatment of the partitioning of synaptic
a All synaptic delay is axonal, backpropagation delay is 0. The synaptic delays
raw cross-correlation function is shifted to the left by d. b The axonal

Bi and Poo 1998). In fact, when a presynaptic spike is gene- synaptic delay, even when individual spike trains are irregular
f
rated at t j , it must first travel down the axon before arri- (see Kriener et al. 2008, for discussion). If the axonal delay is
f
ving at the synapse, thus arriving at t sj = t j + dA , where the same as the backpropagation delay, i.e. dA = dBP = d/2,
dA is the axonal propagation delay. Similarly, a postsynaptic where d is the total transmission delay of the spike, the raw
f
spike at ti must backpropagate through the dendrite before cross-correlation function at the synapse is the same as the
f raw cross-correlation at the soma:
arriving at the synapse at tis = ti + dBP , where dBP is the
backpropagation delay. Consequently, the relevant temporal


 sji t s =  ji t s + (dA − dBP ) =  ji (t) .
difference for STDP update rules is t s = tis − t sj as ini-
tially suggested by Gerstner et al. (1993) and Debanne et al. This situation is depicted in Fig. 8b. Let w0 be the synap-
(1998). Senn et al. (2002) showed that under fairly general tic weight for which the synaptic drift given in (21) is 0,
conditions, STDP may cause adaptation in the presynaptic i.e. the fixed point of the synaptic dynamics for the cross-
and postsynaptic delays in order to optimize the effect of correlation shown. If the axonal delay is larger than the back-
the presynaptic spike on the postsynaptic neuron. In order propagation delay, this results in a shift of the raw cross-
to calculate the synaptic drift as in (17), we therefore need correlation function to the left. This is shown in Fig. 8a for
to integrate the synaptic weight changes over t s , weighted the extreme case of dA = d, dBP = 0, resulting in a net shift
by the raw cross-correlation function at the synapse. With of d. This increases the value of the first integral in (21) and
 ji (t) =  ji (t s + (dA − dBP )), we reformulate (17) as: decreases the second integral, such that ẇ < 0 at w0 . Conver-
sely, if the axonal delay is smaller than the backpropagation
0


 delay, the raw cross-correlation function is shifted to the right
ẇ = −F− (w) dt s K − t s  ji t s + (dA − dBP ) (Fig. 8c, for the extreme case of dA = 0, dBP = d). This
−∞ decreases the value of the first integral in (21) and increases
∞ the second integral, such that ẇ > 0 at w0 . Therefore, a



+F+ (w) dt s K + t s  ji t s + (dA − dBP ) . given network dynamics may cause systematic depression,
0 systematic potentiation or no systematic change at all to the
(21) synaptic weights, depending on the partition of the synap-
tic delay into axonal and dendritic contributions. Systematic
In the case of independent Poisson processes as in synaptic weight changes can in turn result in qualitatively
Sect. 4.1.1.1, the shift of the raw cross-correlation function different network behavior. For example, in Morrison et al.
by (dA − dBP ) has no effect, as  ji (t) is constant. Gene- (2007) small systematic biases in the synaptic weight dyna-
rally, however, this is not the case. For example, networks of mics were applied to a network with an equilibrium charac-
neurons, both in experiment and simulation, typically exhi- terized by a unimodal weight distribution and medium rate
bit oscillations with a period several times larger than the (< 10 Hz) asynchronous irregular activity dynamics. Here, a

123
472 Biol Cybern (2008) 98:459–478

small systematic depression led to a lower weight, lower rate

equilibrium also in the asynchronous irregular regime, whe-
reas a systematic potentiation led to a sudden transition out
of the asynchronous irregular regime: the activity was cha-
racterized by strongly patterned high-rate peaks of activity
interspersed with silence, and the unimodal weight distribu-
tion splintered into several peaks.

4.2 Beyond pair effects

There is considerable evidence that the pair-based rules dis-

cussed above cannot give a full account of STDP. Specifi-
cally, they reproduce neither the dependence of plasticity on
the repetition frequency of pairs of spikes in an experimen-
tal protocol, nor the results of recent triplet and quadruplet
experiments.
STDP experiments are usually carried out with about 60 Fig. 9 Implementation of the triplet rule by local variables. The spikes
of presynaptic neuron j contribute to a trace x j (t), the spikes of postsy-
pairs of spikes. The temporal distance of the spikes in the
naptic neuron i contribute to a fast trace yi1 (t) and a slow trace yi2 (t). The
pair is of the order of a few to tens of milliseconds, whereas update of the weight wi j at the moment of a presynaptic spike is propor-
the temporal distance between the pairs is of the order of tional to the momentary value of the fast trace yi1 (t) (unfilled circles),
hundreds of milliseconds to seconds. In the case of a faci- as in the pair-based model (see Fig. 3). The update of the weight wi j at
the moment of a postsynaptic spike is proportional to the momentary
litation protocol (i.e. pre-before-post), standard pair-based
value of the trace x j (t) (black filled circles) and the value of the slow
STDP models predict that if the repetition frequency is trace yi2 (t) just before the spike (gray filled circles)
increased, the strength of the depressing interaction (i.e. post-
before-pre) becomes greater, leading to less net potentiation.
This prediction is independent of whether the spike pairing dx j xj   f

=− + δ t − tj ,
scheme is all-to-all or nearest neighbor (see Sect. 4.1.2). dt τx f
tj
However, experiments show that increasing the repetition fre-
quency leads to an increase in potentiation (Sjostrom et al. f
2001). Other recent experiments employed multiple-spike where t j denotes the firing times of the presynaptic neuron.
protocols, such as repeated presentations of symmetric tri- Unlike pair-based rules, each spike from postsynaptic neuron
plets of the form pre-post-pre and post-pre-post (Bi and Wang i contributes to a fast trace yi1 and a slow trace yi2 at the
2002; Froemke and Dan 2002; Wang et al. 2005; Froemke synapse:
et al. 2006). Standard pair-based models predict that the two  
dyi1 y1 f
sequences should give essentially the same results, as they = − i + δ t − ti
each contain one pre-post pair and one post-pre pair. Expe- dt τ1
dyi2 y2  
rimentally, quite different results are observed. f
= − i + δ t − ti ,
Here we review two examples of simple models which dt τ2
account for these experimental findings. For other models
where τ1 < τ2 , see Fig. 9. LTD is induced as in the standard
which also reproduce frequency dependence or multiple-
STDP pair model given in (13), i.e. the weight change is
spike protocol results, see Abarbanel et al. (2002), Senn
proportional to the value of the fast postsynaptic trace yi1
(2002) and Appleby and Elliott (2005).
evaluated at the moment of a presynaptic spike. The new
feature of the rule is that LTP is induced by a triplet effect: the
4.2.1 Triplet model
weight change is proportional to the value of the presynaptic
trace x j evaluated at the moment of a postsynaptic spike
One simple approach to modeling STDP which addresses
and also to the slow postsynaptic trace yi2 remaining from
these issues is the triplet rule developed by Pfister and
previous postsynaptic spikes:
Gerstner (2006). This model is based on sets of three spikes
(one presynaptic and two postsynaptic). As in the case of  
  f   f −
wi+j ti = F+ wi j x j ti yi2 ti
f
pair-based rules, the triplet rule can be easily implemented
with local variables as follows. Similarly to pair-based rules,
f−
each spike from presynaptic neuron j contributes to a trace where ti indicates that the function yi2 is to be evaluated
f
x j at the synapse: before it is incremented due to the postsynaptic spike at ti .

123
Biol Cybern (2008) 98:459–478 473

∆t exact relationship between the above triplet model and other

models is discussed in Pfister and Gerstner (2006).
1/ρ
4.2.2 Suppression model

1 An alternative model to address the inability of standard pair-

Experiments, ∆ t = 10 ms
0.8 Experiments, ∆ t = −10 ms
based models to account for data obtained from triplet and
All−to−All quadruplet spike protocols was developed by Froemke and
0.6 Nearest−Spike Dan (2002). They observed that in triplet protocols of the
form pre-post-pre, as long as the intervals between the spikes
0.4
were reasonably short (< 15 ms), the timing of the pre–post
0.2 pair was a better predictor for the change in the synaptic
strength than either the timing of the post–pre pair or of both
0
timings taken together. Similarly, in post–pre–post protocols,
−0.2 the timing of the first post-pre pairing was the best predic-
tor for the change of synaptic strength. On the basis of this
−0.4
observation, they proposed a model in which the synaptic
−0.6 weight change is not just dependent on the timing of a spike
0 10 20 30 40 50 pair, but also on the efficacy of the spikes. Each spike of pre-
synaptic neuron j sets the presynaptic spike efficacy j to 0
Fig. 10 The triplet rule reproduces the finding that increased frequency
whereafter it recovers exponentially to 1 with a time constant
of pair repetition leads to increased potentiation in visual cortex pyra- τ j . The efficacy of the nth presynaptic spike is given by:
midal neurons. Data from Sjostrom et al. (2001), figure adapted from
Pfister and Gerstner (2006)    
n
j = 1 − exp − t nj − t n−1
j /τ j ,

where t nj denotes the nth spike of neuron j. In other words,

Analogously to pair-based models, the triplet rule can also the efficacy of a spike is suppressed by the proximity of a
be implemented with nearest-neighbor rather than all-to-all previous spike. Similarly, the postsynaptic spike efficacy is
spike pairings by an appropriate choice of trace dynamics, reset to 0 by each spike of postsynaptic neuron i, recovering
see Sect. 4.1.2. exponentially to 1 with time constant τi . The model can be
The triplet rule reproduces experimental data from visual implemented with local variables as follows. Each presynap-
cortical slices (Sjostrom et al. 2001) that increasing the repe- tic spike contributes to an efficacy trace j (t) with dynamics:
tition frequency in the STDP pairing protocol increases net
potentiation (Fig. 10). It also gives a good fit to experiments   
d j j −1 − f
based on triplet protocols in hippocampal culture (Wang et al. =− − j δ t − tj ,
dt τj f
2005). The main functional advantage of such a triplet lear- tj
ning rule is that it can be mapped to a Bienenstock–Cooper–
Munro learning rule (Bienenstock et al. 1982): if we assume where − j denotes the value of j just before the update. The
that the pre- and postsynaptic spike trains are governed by standard presynaptic trace x j given in (11) is adapted to take
Poisson statistics, the triplet rule exhibits depression for low the spike efficacy into account:
postsynaptic firing rates and potentiation for high postsy-
naptic firing rates. If we further assume that the triplet term dx j xj   
− f
in the learning rule depends on the mean postsynaptic fre- =− + j δ t − tj ,
dt τx
quency, a sliding threshold between potentiation and depres- f
tj
sion can be defined. In this way, the learning rule matches
the requirements of the BCM theory and inherits the pro- i.e. each presynaptic spike increments x j by the value of the
perties of the BCM learning rule such as input selectivity. spike efficacy before the update. Similarly, each postsynaptic
From BCM properties, we can immediately conclude that spike contributes to an efficacy trace i (t) with dynamics:
the model should be useful for receptive field development.
  
Note that earlier efforts to show that STDP maps to the BCM d i i −1 − f
model (Izhikevich and Desai 2003; Senn et al. 2000) demons- =− − i δ t − ti ,
dt τi f
trated neither an exact mapping nor a sliding threshold. The ti

123
474 Biol Cybern (2008) 98:459–478

and a postsynaptic trace yi with increments weighted by the depolarization in the absence of postsynaptic firing cannot
postsynaptic spike efficacy: be modeled by standard STDP or triplet models.
dyi yi   
− f
=− + δ t − t . 4.4 Induction versus maintenance
dt τy f
i i
ti
We stress that all the above models concern induction of
The weight updates on the occurrence of a post- or presy-
potentiation and depression, but not their maintenance. The
naptic spike are therefore given by:
induction of LTP may take only a few seconds: for example,
 
  f  − f 
wi+j ti = F+ wi j x j ti stimulation with 50 pairs of pre- and postsynaptic spikes
f
i ti
      given at 20 Hz takes less than 3 s. However, afterwards the


wi−j t j = −F− wi j yi t j − f
f f
j tj .
synapse takes 60 min or more to consolidate these changes,
and this process may also be interrupted (Frey and Morris
This model gives a good fit to triplet and quadruplet protocols 1997). During this time synapses are ‘tagged’, that is, they
in visual cortex slice, and also gives a much better prediction are ready for consolidation. Consolidation is thought to rely
for synaptic modification due to natural spike trains (Froemke on a different molecular mechanism than that of induction.
and Dan 2002). However, it does not predict the increase Simply speaking, gene transcription is necessary to trigger
of LTP with the repetition frequency observed by Sjostrom the building of new proteins that increase the synaptic effi-
et al. (2001). A revised version of the model (Froemke et al. cacy.
2006) also accounts for the switch of LTD to LTP at high
frequencies by modifying the efficacy functions. 4.4.1 Functional consequences

4.3 Voltage dependence Long-term stability of synapses is necessary to retain memo-

ries that have been learned, despite ongoing activity of presy-
Traditional LTP/LTD experiments employ the following naptic neurons. A simple possibility used in many models is
induction paradigm: the postsynaptic neuron is held at a that plasticity is simply switched off once the neuron has lear-
fixed depolarization while one or several presynaptic neu- ned what it should. This approach makes sense in the context
rons are activated. Often a presynaptic pathway is stimula- of reward-based learning: the learning rate goes to zero once
ted extracellularly, so that several presynaptic neurons are the actual reward equals the expected reward and learning
activated. Depending on the level of the postsynaptic mem- stops automatically (see Sect. 5.2). It also makes sense in the
brane potential, the activated synapses increase their effi- framework of supervised learning (see Sect. 5.1). Learning is
cacy while other non-activated synapses do not change their normally driven by the difference between desired output and
weight (Artola et al. 1990; Artola and Singer 1993). More actual output. However, in the context of unsupervised lear-
recently, depolarization has also been combined with STDP ning it is inconsistent to switch off the dynamics. Neverthe-
experiments. In particular, Sjostrom et al. (2004) showed a less, receptive field properties should be retained for a fairly
dependence of synaptic weight changes on the synaptic mem- long time even if the stimulation characteristic changes.
brane potential just before a postsynaptic spike.
There is an ongoing discussion whether the voltage 4.4.2 Bistability model
dependence is more fundamental than the dependence on
postsynaptic spiking. Indeed, voltage dependence alone can A simple model of maintenance has been proposed by Fusi
generate STDP-like behavior (Brader et al. 2007), as the et al. (2000). The basis of the model is a hidden variable that
membrane potential behaves in a characteristic way in the has an unstable fixed point (threshold). If the variable has
vicinity of a spike (high shortly before a spike, and low a value above threshold it converges towards 1; otherwise
shortly after). Alternatively, a dependence on the slope of towards 0. To stay within the framework of the previous sec-
the postsynaptic membrane potential has also been shown tions, let us suppose that the weight w is calculated by one
to reproduce the characteristic STDP weight change curve of the STDP or short-term plasticity models. Maintenance is
(Saudargiene et al. 2003). The voltage effects caused by back- implemented by adding on top of the STDP dynamics a slow
propagating spikes is implicitly contained in the mechanis- bistable dynamics (Gerstner and Kistler 2002):
tic formulation of STDP models outlined above. In particu-
lar, the fast postsynaptic trace y1 in the above triplet model dw
= α(w) = −w (1 − w)(wθ − w)/τα ,
could be seen as an approximation of a backpropagating dt
action potential. However, the converse is not true: a pure where τα is a time constant of consolidation in the range
STDP rule does not automatically generate a voltage depen- of several minutes of biological time. The result is that in
dence. Moreover, synaptic effects caused by subthreshold the absence of any stimulation, individual synapses evolve

123
Biol Cybern (2008) 98:459–478 475

towards binary values of 0 or 1 which are intrinsically stable Theoretical studies have demonstrated that a teacher-forced
fixed points of the slow dynamics. As a result, rather strong STDP approach can be used to learn precise spike times
stimuli are necessary to perturb the synaptic dynamics. (Legenstein et al. 2005; Pfister et al. 2006). In a natural
situation, this would mean that a few strong neural inputs
4.4.3 Biological evidence can drive the neuron and therefore drive learning of other
inputs. If these strong inputs are controlled in a task-specific
Whether single synapses themselves are binary or continuous way, they act as a teacher for the postsynaptic neuron. For a
is a matter of intense debate. Some experiments have sugges- practical realization of this idea see Brader et al. (2007).
ted that synapses are binary (Petersen et al. 1998; O’Connor
et al. 2005). However, this would seem to result in a bistable 5.2 Reinforcement learning
distribution of weights which is at odds with the unimodal
distribution reported by other studies (Turrigiano et al. 1998; If neuronal activity leads to actions, feedback may arise from
Sjostrom et al. 2001; Song et al. 2005), and with the finding the environment in forms of reward (a piece of pizza) or
that the magnitude of LTP/LTD increases with the number of punishment (burnt fingers). It is thought that success of an
spike pairs in a protocol until saturation is reached (Froemke action is signaled by neuromodulators—a top candidate is
et al. 2006). dopamine (Schultz et al. 1997). Dopamine signals are clo-
Some possibilities to reconcile these findings include: (i) sely related to a quantity in reinforcement learning known as
since pairs of neurons form several contacts with each other, δ, that can be interpreted as the difference between the recei-
it is likely that in standard plasticity experiments several ved reward and the expected reward. Here ‘reward’ means
synapses are measured at the same time; (ii) LTP and STDP current or future rewards that can be reliably predicted. In
results are typically reported as pooled experiments over reinforcement learning, the difference between actual and
several pairs of neurons. Under the assumption that the upper expected rewards plays an important role for the update of
bound is not the same for all synapses, a broad distribution weights in Q-learning, SARSA, and related variants of tem-
could result; (iii) both unimodal distribution and bimodal dis- poral difference learning (Sutton and Barto 1998).
tributions could be stable. Untrained neurons would show a Under a suitable interpretation of the role of pre- and post-
unimodal distribution whereas neurons that have learned to synaptic neurons, the weight update rules can be derived from
respond to a specific pattern would develop a bimodal dis- an optimality framework (Pfister et al. 2006). The learning
tribution of synaptic weights (Toyoizumi et al. 2007); (iv) rule can be interpreted as a Hebbian learning based on joined
all synapses are binary, but the efficacy of the ‘strong’ state pre- and postsynaptic activity, but conditioned on the pre-
is subject to short-term plasticity and homeostasis; (v) some sence of a global reward signal. Variants of such reinforce-
synapses are binary and some are not. Potentially a combi- ment rules for spiking neurons have been developed (Seung
nation of several of these possibilities must be considered in 2003; Pfister et al. 2006; Izhikevich 2007; Florian 2007).
order to explain the experimental findings.

6 Discussion
5 Supervised and reinforcement learning
Pair-based STDP models can be decomposed into three
All the models considered in Sect. 4 are unsupervised ‘Heb- aspects: weight dependence, spike-pairing scheme and delay
bian’ rules: changes are triggered as a result of combined partition (Sect. 4.1). We have shown that all of these aspects
action of pre- and postsynaptic neurons. The postsynaptic can have significant consequences for the behavior of the
neuron itself is driven by its input arising from presynaptic model system under investigation. However, in many cases
neurons. There is no notion of whether or not the postsynaptic there is not enough experimental data to settle these ques-
output is ‘good’ or ‘useful’. If, however, the local variables tions definitively. Therefore, choices for each aspect should
are combined with global teacher or reinforcement signals, be made consciously and take into consideration the relevant
completely different learning paradigms are possible. available experimental findings. Moreover, these choices
should be explicitly documented and critically addressed: it
5.1 Supervised learning should be clear to what extent results depend on the specific
choices.
Supervised plasticity has been demonstrated experimentally In particular, the choice of STDP weight dependence is
by Fregnac and Schulz (2006): the behavior of a (cortical) critical. The available evidence suggests that both potentia-
neuron can be changed by pairing some class of stimuli tion and depression are dependent on the weight. Whereas it
with an (artificial) increase of neural activity while pairing is useful to start with very simplified models to gain insight,
another class of stimuli with a decrease of responsiveness. we now know that STDP models which assume some weight

123
476 Biol Cybern (2008) 98:459–478

dependence produce qualitatively different behavior from the tation in NEST, and G. Hennequin for proofreading the manuscript.
additive model. Moreover, weight dependent rules are no har- We are very grateful to G-q. Bi and M-m. Poo for providing us with
their original data. This work was partially funded by EU Grant 15879
der to implement computationally than additive rules. In the (FACETS), DIP F1.2 and BMBF Grant 01GQ0420 to the Bernstein
absence of fresh experimental evidence supporting an addi- Center for Computational Neuroscience Freiburg.
tive rule, weight dependent rules should therefore be consi-
dered as the standard. Open Access This article is distributed under the terms of the Creative
Commons Attribution Noncommercial License which permits any
Pair-based models of STDP have their limitations. They noncommercial use, distribution, and reproduction in any medium,
give incorrect predictions for many experiments such as tri- provided the original author(s) and source are credited.
plet and quadruplet protocols and cannot account for synaptic
modification due to natural spike trains or pairing protocols
References
at different frequencies. Models of STDP that are beyond
the pair-based framework (Sect. 4.2) can account for these Abarbanel H, Huerta R, Rabinovich M (2002) Dynamical model of
findings at the cost of only a small number of additional long-term synaptic plasticity. Proc Natl Acad Sci USA 99(15):
variables, and so should attract increasing theoretical inter- 10132–10137
Abbott LF, Nelson SB (2000) Synaptic plasticity: taming the beast. Nat
est.
Neurosci 3(Suppl):1178–1183
In this manuscript, we have considered models in which Abbott LF, Varela JA, Sen K, Nelson SB (1997) Synaptic depression
synaptic modifications depend only on spike timing. Howe- and cortical gain control. Science 275:220–223
ver, this ignores many aspects of synaptic plasticity which Appleby P, Elliott T (2005) Synaptic and temporal ensemble interpreta-
tion of spike-timing-dependent plasticity. Neural Comput 17(11):
may prove to be of great importance to the functioning of the
2316–2336
brain, and will therefore have to be taken into consideration Artola A, Bröcher S, Singer W (1990) Different voltage dependent thre-
in future phenomenological modeling. Most STDP models sholds for inducing long-term depression and long-term potentia-
assume that the absolute synaptic strength is modified (but tion in slices of rat visual cortex. Nature 347:69–72
Artola A, Singer W (1993) Long-term depression of excitatory synaptic
see Senn 2002). However, it may turn out that a formulation
transmission and its relationship to long-term potentiation. Trends
in terms of the release probability is a more accurate descrip- Neurosci 16(11):480–487
tion, thus allowing a unified view of short-term and long-term Badoual M, Zou Q, Davison AP, Rudolph M, Bal T, Fregnac Y,
plasticity. Additionally, STDP has been shown to be sensi- Destexhe A (2006) Biophysical and phenomenological models of
multiple spike interactions in spike-timing dependent plasticity.
tive to a number of factors beyond spike timing, for example
Int J Neural Systems 16:79–97
active dendritic properties and the location of the synapse on Bell C, Han V, Sugawara Y, Grant K (1997) Synaptic plasticity in
the dendrite - see Kampa et al. (2007) for a review. There is a cerebellum-like structure depends on temporal order. Nature
also substantial evidence that inhibition is an important phy- 387:278–281
Bi G-q, Poo M (2001) Synaptic modification by correlated activity:
siological feature fine-tuning induction and maintenance of
Hebb’s postulate revisited. Annu Rev Neurosci 24:139–166
LTP/LTD. Inhibition gates induction of LTP/LTD as a func- Bi G-q, Wang H (2002) Temporal asymmetry in spike timing-
tion of physiological conditions and physiologically-induced dependent synaptic plasticity. Physiol Behav 77:551–555
changes in the activity of networks (Larson and Lynch 1986; Bi G-q, Poo M-m (1998) Synaptic modifications in cultured hippocam-
pal neurons: Dependence on spike timing, synaptic strength, and
Pacelli et al. 1989; Radpour and Thomson 1991; Steele and
postsynaptic cell type. J Neurosci 18:10464–10472
Mauk 1999; Nishiyama et al. 2000; Togashi et al. 2003). Bienenstock EL, Cooper LN, Munro PW (1982) Theory for the deve-
Here, the main challenge is to derive appropriate phenome- lopment of neuron selectivity: orientation specificity and binocular
nological models from experiments and detailed biophysi- interaction in visual cortex. J Neurosci 2(1):32–48
Billings G, van Rossum M (2008) Memory retention and spike timing
cal models. Finally, although some progress has been made
dependent plasticity (preprint)
in investigating the interactions of STDP with other plas- Bliss TVP, Collingridge GL (1993) A synaptic model of memory: long-
ticity mechanisms such as homeostasis and heterosynaptic term potentiation in the hippocampus. Nature 361:31–39
spread of LTP/LTD (van Rossum et al. 2000; Toyoizumi et al. Bliss TVP, Lomo T (1973) Long-lasting potentation of synaptic trans-
mission in the dendate area of anaesthetized rabbit following sti-
2005, 2007; Triesch 2007), this complex topic remains lar-
mulation of the perforant path. J Physiol 232:331–356
gely unexplored. In this area, the main challenge is to perform Brader JM, Senn W, Fusi S (2007) Learning real world stimuli in a neu-
analytical and simulation studies which can identify and cha- ral network with spike-driven synaptic dynamics. Neural Comput
racterize their composite effects, and investigate their func- 19(11):2881–2912
Burkitt AN, Gilson M, van Hemmen JL (2007) Spike-timing-dependent
tional consequences.
plasticity for neurons with recurrent connections. Biol Cybern
96(5):533–546
Acknowledgements The idea for this paper grew out of a FACETS Burkitt AN, Meffin H, Grayden DB (2004) Spike-timing-dependent
workshop on synaptic plasticity held in Lausanne in June 2006. We plasticity: the relationship to rate-based learning for models with
would therefore like to thank all the participants, especially A. Davi- weight dynamics determined by a stable fixed point. Neural
son, A. Destexhe, Y. Fregnac, C. Lamy, R. Legenstein, W. Maass, J.-P. Comput 16:885–940
Pfister, and A. Thomson for their contributions. We also thank M. Helias Cooper L, Intrator N, Blais B, Shouval HZ (2004) Theory of cortical
for helpful discussions about short-term plasticity and the implemen- plasticity. World Scientific, Singapore

123
Biol Cybern (2008) 98:459–478 477

Dayan P, Abbott LF (2001) Theoretical neuroscience. MIT Press, Kandel ER, Schwartz JH, Jessel TM (2000) Principles of neural science,
Cambridge 4th edn. McGraw-Hill, New York. ISBN 978-0838577011
Debanne D, Gähwiler BH, Thompson SM (1998) Long-term synaptic Kelso SR, Ganong AH, Brown TH (1986) Hebbian synapses in hippo-
plasticity between pairs of individual CA3 pyramidal cells in rat campus. Proc Natl Acad Sci USA 83:5326–5330
hippocampal slice cultures. J Physiol (Lond) 507:237–247 Kempter R, Gerstner W, van Hemmen JL (1999) Hebbian learning and
Dudek SM, Bear MF (1992) Homosynaptic long-term depression in spiking neurons. Phys Rev E 59:4498–4514
area CA1 of hippocampus and effects of n-methyl-d-aspartate Kempter R, Gerstner W, van Hemmen JL (2001) Intrinsic stabilization
receptor blockade. Proc Natl Acad Sci USA 89:4363–4367 of output rates by spike-based Hebbian learning. Neural Comput
Florian RV (2007) Reinforcement learning through modulation of 12:2709–2742
spike-timing-dependent synaptic plasticity. Neural Comput Kistler WM, van Hemmen JL (2000) Modeling synaptic plasticity in
19:1468–1502 conjunction with the timing of pre- and postsynaptic action poten-
Fregnac Y, Schulz DE (2006) Activity-dependent regulation of recep- tials. Neural Comput 12:385–405
tive field properties of cat area 17 by supervised Hebbian learning. Kriener B, Tetzlaff T, Aertsen A, Diesmann M, Rotter S (2008) Cor-
J Neurobio 41:69–82 relations and population dynamics in recurrent cortical networks.
Frey U, Morris R (1997) Synaptic tagging and long-term potentiation. Neural Comput (in press)
Nature 385:533–536 Larson J, Lynch G (1986) Induction of synaptic potentiation in hip-
Froemke R, Tsay I, Raad M, Long J, Dan Y (2006) Contribution of indi- pocampus by patterned stimulation involves two events. Science
vidual spikes in burst-induced long-term synaptic modification. 232:985–988
J Neurophysiol 95:1620–1629 Legenstein R, Naeger C, Maass W (2005) What can a neuron learn with
Froemke RC, Dan Y (2002) Spike-timing-dependent synaptic modifi- spike-timing-dependent plasticity? Neural Comput 17(11):2337–
cation induced by natural spike trains. Nature 416(6879):433–438 2382
Fusi S, Annunziato M, Badoni D, Salamon A, Amit DJ (2000) Spike- Levy WB, Steward D (1983) Temporal contiguity requirements for
driven synaptic plasticity: Theory, simulation, VLSI implementa- long-term associative potentiation/depression in the hippocampus.
tion. Neural Comput 12(10):2227–2258 Neuroscience 8:791–797
Fusi S, Drew PJ, Abbott LF (2005) Cascade models of synaptically Lisman J (1989) A mechanism for Hebb and anti-Hebb processes
stored memories. Neuron 45(4):599–611 underlying learning and memory. Proc Natl Acad Sci USA 86:
Gerstner W, Kempter R, van Hemmen JL, Wagner H (1996) A neuronal 9574–9578
learning rule for sub-millisecond temporal coding. Nature 383: Lisman JE, Zhabotinsky AM (2001) A model of synaptic memory: A
76–78 CaMKII/PP1 switch that potentiates transmission by organizing
Gerstner W, Kistler W (2002) Spiking neuron models: single neurons, an AMPA receptor anchoring assembly. Neuron 31:191–201
populations, plasticity. Cambridge University Press, Cambridge Lu J, Li C, Zhao JP, Poo M-m, Zhang X (2007) Spike-timing-dependent
Gerstner W, Ritz R, van Hemmen JL (1993) Why spikes? Hebbian lear- plasticity of neocortical excitatory synapses on inhibitory interneu-
ning and retrieval of time-resolved excitation patterns. Biol Cybern rons depends on target cell type. J Neurosci 27:9711–9720
69(5–6):503–515 Malenka RC, Kauer J, Zucker R, Nicoll RA (1988) Postsynaptic cal-
Gewaltig M-O, Diesmann M (2007) NEST (neural simulation tool). cium is sufficient for potentiation of hippocampal synaptic trans-
Scholarpedia 2(4):1430 mission. Science 242:81–84
Graupner M, Brunel N (2007) STDP in a bistable synapse model based Malenka RC, Nicoll RA (1993) NMDA-receptor-dependent plasticity:
on CaMKII and associated signaling pathways. Public Library Sci multiple forms and mechanisms. Trends Neurosci 16:480–487
Comput Biol 3(11):e221 Markram H, Lübke J, Frotscher M, Sakmann B (1997) Regulation of
Grossberg S (1987) The adaptive brain I. Elsevier, Amsterdam synaptic efficacy by coincidence of postsynaptic APs and EPSPs.
Gupta A, Wang Y, Markram H (2000) Organizing principles for a diver- Science 275:213–215
sity of GABAergic interneurons and synapses in the neocortex. Markram H, Sakmann B (1995) Action potentials propagating back into
Science 287:273–278 dendrites trigger changes in efficacy of single-axon synapses bet-
Gustafsson B, Wigstrom H, Abraham WC, Huang Y-Y (1987) Long- ween layer V pyramidal neurons. Soc Neurosci Abstr 21(3):2007
term potentiation in the hippocampus using depolarizing current Markram H, Wang Y, Tsodyks M (1998) Differential signaling via the
pulses as the conditioning stimulus to single volley synaptic poten- same axon of neocortical pyramidal neurons. Proc Natl Acad Sci
tials. J Neurosci 7:774–780 USA 95(9):5323–5328
Gütig R, Aharonov R, Rotter S, Sompolinsky H (2003) Learning input Miller K, Keller JB, Stryker MP (1989) Ocular dominance column
correlations through nonlinear temporally asymmetric Hebbian development: analysis and simulation. Science 245:605–615
plasticity. J Neurosci 23(9):3697–3714 Morrison A, Aertsen A, Diesmann M (2007) Spike-timing dependent
Hebb DO (1949) The organization of behavior: a neuropsychological plasticity in balanced random networks. Neural Comput 19:1437–
theory. Wiley, New York 1467
Iglesias J, Eriksson J, Grize F, Tomassini M, Villa A (2005) Dynamics Morrison A, Straube S, Plesser HE, Diesmann M (2007) Exact sub-
of pruning in simulated large-scale spiking neural networks. Bio- threshold integration with continuous spike times in discrete time
systems 79:11–20 neural network simulations. Neural Comput 19(1):47–79
Izhikevich EM (2007) Solving the distal reward problem through lin- Ngezahayo A, Schachner M, Artola A (2000) Synaptic activity modu-
kage of STDP and dopamine signaling. Cereb Cortex 17(10): lates the induction of bidirectional synaptic changes in adult mouse
2443–2452 hippocampus. J Neurosci 20(7):2451–2458
Izhikevich EM, Desai NS (2003) Relating STDP to BCM. Neural Nishiyama M, Hong K, Mikoshiba K, Poo M, Kato K (2000) Calcium
Comput 15:1511–1523 stores regulate the polarity and input specificity of synaptic modi-
Izhikevich EM, Gally JA, Edelman GM (2004) Spike-timing dynamics fication. Nature 408(6812):584–588
of neuronal groups. Cereb Cortex 14:933–944 O’Connor D, Wittenberg G, Wang S-H (2005) Graded bidirectional
James W (1890) Psychology (briefer course). Holt, New York synaptic plasticity is composed of switch-like unitary events. Proc
Kampa BM, Letzkus JJ, Stuart GJ (2007) Dendritic mechanisms Natl Acad Sci USA 102:9679–9684
controlling spike-timing-dependent synaptic plasticity. Trends Oja E (1982) A simplified neuron model as a principal component ana-
Neurosci 30(9):456–463 lyzer. J Math Biol 15:267–273

123
478 Biol Cybern (2008) 98:459–478

Pacelli GJ, Sue W, Keslo SR (1989) Activity-induced depression of Song S, Per S, Reigl M, Nelson S, Chklovskii D (2005) Highly non-
synaptic inhibition during LTP-inducing patterned stimulation. random features of synaptic connectivity in local cortical circuits.
Brain Res 486:26–32 Public Library Sci Biol 3(3):0507–0519
Petersen C, Malenka R, Nicoll R, Hopfield J (1998) All-or-none poten- Standage D, Jalil S, Trappenberg T (2007) Computational conse-
tiation of CA3-CA1 synapses. Proc Natl Acad Sci USA 95:4732– quences of experimentally derived spike-time and weight
4737 dependent plasticity rules. Biol Cybern 96(6):615–623
Pfister J-P, Gerstner W (2006) Triplets of spikes in a model of spike Steele PM, Mauk MD (1999) Inhibitory control of LTP and LTD:
timing-dependent plasticity. J Neurosci 26:9673–9682 stability of synapse strength. J Neurophysiol 81:1559–1566
Pfister J-P, Toyoizumi T, Barber D, Gerstner W (2006) Optimal spike- Sutton RS, Barto AG (1998) Reinforcement learning: an introduction.
timing dependent plasticity for precise action potential firing in Adaptive Computation and Machine Learning. The MIT Press
supervised learning. Neural Comput 18:1309–1339 Thomson AM, Deuchars J, West DC (1993) Large, deep layer pyramid-
Radpour S, Thomson AM (1991) Coactivation of local circuit NMDA pyramid single axon EPSPs in slices of rat motor cortex display
receptor mediated EPSPs induces lasting enhancement of mini- paired pulse and freuquency-dependent depression, mediated
mal schaffer collateral epsps in slices of rat hippocampus. Eur J presynaptically and self-facilitation mediated postsynaptically.
Neurosci 3:602–613 J Neurophysiol 70(6):2354–2369
Roberts PD (1999) Computational consequences of temporally asym- Thomson AM, Lamy C (2007) Functional maps of neocortical local
metric learning rules: I. Differential Hebbian learning. J Comput circuitry. Front Neurosci 1:19–42
Neurosci 7:235–246 Togashi K, Kitajima T, Aihara T, Hong K, Poo M, Nishiyama M (2003)
Rotter S, Diesmann M (1999) Exact digital simulation of time-invariant Gating of activity-dependent long-term depression by GABAergic
linear systems with applications to neuronal modeling. Biol activity in the hippocampus. In: SocNeurosciAbstr, pp 123.4
Cybern 81(5/6):381–402 Toyoizumi T, Pfister J, Aihara K, Gerstner W (2007) Optimality
Rubin J, Lee D, Sompolinsky H (2001) Equilibrium properties of tem- model of unsupervised spike-timing-dependent plasticity: synaptic
porally asymmetric Hebbian plasticity. Phys Rev Lett 86:364–367 memory and weight distribution. Neural Comput 19(3):639–671
Rubin JE, Gerkin RC, Bi G-q, Chow CC (2005) Calcium time course Toyoizumi T, Pfister JP, Aihara K, Gerstner W (2005) Generalized
as a signal for spike-timing-dependent plasticity. J Neurophysiol Bienenstock-Cooper-Munro rule for spiking neurons that maxi-
93:2600–2613 mizes information transmission. Proc Natl Acad Sci USA 102(14):
Saudargiene A, Porr B, Wörgötter F (2003) How the shape of pre- and 5239–5244
postsynaptic signals can influence STDP: a biophysical model. Triesch J (2007) Synergies between intrinsic and synaptic plasticity
Neural Comput 16:595–626 mechanisms. Neural Comput 19(4):885–909
Schemmel J, Gruebl A, Meier K, Mueller E (2006) Implementing Tsodyks M, Pawelzik K, Markram H (1998) Neural networks with
synaptic plasticity in a VLSI spiking neural network model. In: dynamic synapses. Neural Comput 10:821–835
Proceedings of the 2006 international joint conference on neural Tsodyks M, Uziel A, Markram H (2000) Synchrony generation in recur-
networks. IEEE Press, pp 1–6 rent networks with frequency-dependent synapses. J Neurosci 20,
Schultz W, Dayan P, Montague PR (1997) A neural substrate of pre- RC1 (1–5)
diction and reward. Science 275:1593–1599 Tsodyks MV, Markram H (1997) The neural code between neocortical
Senn W (2002) Beyond spike timing: the role of nonlinear plasticity pyramidal neurons depends on neurotransmitter release probabi-
and unreliable synapses. Biol Cybern 87:344–355 lity. Proc Natl Acad Sci USA 94:719–723
Senn W, Markram H, Tsodyks M (2000) An algorithm for modifying Turrigiano G, Abbott LF, Marder E (1994) Activity-dependent changes
neurotransmitter release probability based on pre- and postsynaptic in the intrinsic properties of pyramidal neurons. Science 264:974–
spike timing. Neural Comput 13:35–67 977
Senn W, Schneider M, Ruf B (2002) Activity-dependent development Turrigiano GG, Leslie KR, Desai NS, Rutherford LC, Nelson SB
of axonal and dendritic delays, or, why synaptic transmission (1998) Activity-dependent scaling of quantal amplitude in neo-
should be unreliable. Neural Comput 14(3):583–619 cortical neurons. Nature 391:892–896
Seung HS (2003) Learning spiking neural networks by reinforcement Turrigiano GG, Nelson SB (2004) Homeostatic plasticity in the deve-
of stochastic synaptic transmission. Neuron 40:1063–1073 loping nervous system. Nat Rev Neurosci 5:97–107
Shouval HZ, Bear MF, Cooper LN (2002) A unified model of NMDA Rossum MCWvan , Bi G-q, Turrigiano GG (2000) Stable Hebbian lear-
receptor dependent bidirectional synaptic plasticity. Proc Natl ning from spike timing-dependent plasticity. J Neurosci 20(23):
Acad Sci USA 99:10831–10836 8812–8821
Sjostrom P, Turrigiano G, Nelson S (2001) Rate, timing, and cooperati- Wang H-X, Gerkin RC, Nauen DW, Bi G-q (2005) Coactivation and
vity jointly determine cortical synaptic plasticity. Neuron 32:1149– timing-dependent integration of synaptic potentiation and depres-
1164 sion. Nat Neurosci 8(2):187–193
Sjostrom PJ, Turrigiano GG, Nelson SB (2004) Endocannabinoid- Zou Q, Destexhe A (2007) Kinetic models of spike-timing dependent
dependent neocortical layer-5 LTD in the absence of postsynaptic plasticity and their functional consequences in detecting correla-
spiking. J Neurophysiol 92(6):3338–3343 tions. Biol Cybern 97(1):81–97
Song S, Miller KD, Abbott LF (2000) Competitive Hebbian learning
through spike-timing-dependent synaptic plasticity. Nat Neurosci
3(9):919–926

123