Throughout life and development, synaptic inputs are formed at distances of tens to hundreds of micrometers from the soma (Terauchi and Umemori, 2012). Inputs stemming from these synapses are often integrated with one another before they reach the soma and ultimately evoke (or fail to evoke) action potentials. It was recently suggested that dendrites act as independent computational units (Poirazi et al., 2003; Losonczy et al., 2008; Branco and Häusser, 2010) that locally regulate many important cellular processes, such as plasticity and protein synthesis. In this light, a dendritic, local regulation of excitatory and inhibitory synapses should exist and be actively maintained during synaptic plasticity (Liu, 2004; Liu et al., 2007; Bourne and Harris, 2011).

The formation of excitatory synapses is often mediated by the outgrowth of small dendritic protrusions and synapses can thus form between two neurons that previously had no physical interaction. Using two-photon microscopy, Wierenga and colleagues recently showed that inhibitory synapses are formed in a fundamentally different way (Wierenga et al., 2008). New inhibitory synapses grew in locations where an inhibitory axon is already in close contact with a postsynaptic dendrite. The question emerges of what determines the timing and location for the creation of a new inhibitory synapse along the inhibitory axon.

One possibility is that coordinated forms of pre- and postsynaptic activity could play a role in these processes, perhaps even at the level of individual boutons and spines that intersect or are in close proximity to each other. Wierenga studied inhibitory plasticity by monitoring the dynamic behavior of the inhibitory boutons along the axon in hippocampal organotypic cultures of transgenic mice using two-photon microscopy, and showed that even during baseline activity inhibitory boutons are highly dynamic (Schuemann et al., 2013). Roughly 80% of inhibitory boutons were present during the entire 4–5 h imaging period, most likely reflecting stable inhibitory synapses. The other boutons showed highly dynamic behavior. Boutons appeared, disappeared and reappeared at specific locations along the inhibitory axons, presumably axon-dendrite crossings. In addition, these dynamic boutons showed large variance in their size. This indicates active trafficking of presynaptic material, and competition between neighboring boutons along inhibitory axons, similar to what was previously shown for excitatory axons (Staras, 2007). It also suggests that inhibitory axons are continuously exploring potential locations for the formation of new synaptic contacts.

Once a connection between neurons is established, the strength of the synapse remains to be tuned according to its function. In EE connections such tuning often happens through activity-dependent changes, the sign and degree of which crucially depends on the timing of pre- and postsynaptic spiking (Markram et al., 1997; Bi and Poo, 1998; Dan and Poo, 2004). Several studies address the question whether similar rules apply to inhibitory connections (Figure 1). We have separated their results here by the proposed function of the effect.

In a previous study of inhibitory plasticity (although not necessarily iSTDP), Tao and Poo (2005) examined the organization of excitatory and inhibitory spatial receptive fields during development of the Xenopus optic tectum. Initially in young animals (e.g., stage 44 tadpoles), the spatial extent of synaptic receptive fields were quite broad, both for excitation and inhibition; however, excitatory and inhibitory fields were somewhat mismatched and had substantial non-overlapping regions. Over development, synaptic receptive field size was reduced, and excitatory and inhibitory fields became similar, in a manner that seemed to require specific temporal patterns of activity of tectal neurons.

Another example of inhibitory plasticity governing receptive field organization is the self-balancing of excitation and inhibition along sensory processing pathways in auditory cortex. Excitatory-inhibitory balance is a fundamental property of cortical networks, important for control of spike generation, information processing, synaptic plasticity, and prevention of epilepsy (Van Vreeswijk and Sompolinsky, 1996; Moore and Nelson, 1998; Wehr and Zador, 2003; Vogels and Abbott, 2005; Higley and Contreras, 2006; Froemke et al., 2007; de la Rocha et al., 2008; Monier et al., 2008; Okun and Lampl, 2008; Vogels and Abbott, 2009; Cafaro and Rieke, 2010). Recent results in a number of sensory systems indicate that in mature cortex, the strengths and response profiles of inhibitory inputs are in proportion to the strengths and profiles of excitatory inputs (Wehr and Zador, 2003; Higley and Contreras, 2006; Cafaro and Rieke, 2010; House et al., 2011; Saar et al., 2011). Specifically, in adult rat primary auditory cortex (A1), synaptic frequency tuning curves for excitation and inhibition are generally co-tuned and highly correlated [average correlation coefficient r: ≈0.7, Froemke et al. (2007)].

In contrast, in developing A1 of young rats Froemke et al. (2007) reported low correlation between excitation and inhibition. This was not due to a lack of inhibition. Inhibitory responses were present in young animals just after hearing onset, and the overall ratio of excitatory to inhibitory strengths seemed to be conserved in young and adult animals. Rather, the low correlation between excitation and inhibition reflected the untuned nature of inhibitory responses, similar to the observations in developing Xenopus tectum (Tao and Poo, 2005). While excitatory synaptic tuning curves were well-tuned and structured in young animals, inhibitory tuning curves were tuned randomly and generally broader in spectral extent. Changes in the pattern of sensory experience could accelerate the development of excitatory-inhibitory balance, and repetitive presentation of tones of a single frequency led to network-wide plasticity and adjustment of synaptic strength, calibrating excitation and inhibition and shifting the preferred frequency, due to an orchestrated set of long-term excitatory and inhibitory synaptic modifications (Dorrn et al., 2010).

Long-lasting shifts in preferred frequency could also be induced in adult A1, by pairing patterned sensory stimulation with direct activation of neuromodulatory centers. Froemke et al. (2007) focused on the effects of electrical stimulation of the cholinergic basal forebrain, involved in control of selective attention. Muscarinic receptor activation in A1 led to a transient disinhibition, breaking excitatory-inhibitory balance and gating induction of NMDA receptor-dependent synaptic plasticity. Excitatory tuning curves shifted toward the paired tone within ~10 min by two major mechanisms: enhancements of responses at the paired frequency, and decreases in responses at the original preferred frequency. Remarkably, over a longer timescale (several hours), inhibitory tuning curves shifted to re-balance the profiles of excitation, recovering excitatory-inhibitory balance (Froemke et al., 2007).

Haas et al. (2006) investigated spike timing-dependent plasticity of inhibitory synapses (iSTDP) in the entorhinal cortex, a brain area richly associated with spatial navigation (Hafting et al., 2005). Postsynaptic spikes were paired with extracellular stimulations that, in the presence of excitatory synaptic blockade, resulted in inhibitory postsynaptic potentials (IPSPSs). The amplitude of the inhibitory conductance was measured as the slope of the IPSP, before and after spike pairings. For presynaptic inputs preceding postsynaptic spikes, IPSPs were potentiated, with a maximal effect around Δt = −10 ms (Δt = t_pre −t_post) of delay (Figure 1H). In other words, the postsynaptic spike had to arrive 10 ms after the presynaptic inhibitory spike for maximal strengthening. For inhibitory inputs that followed after the postsynaptic spike, a depression of synaptic strength was observed, with a peak effect near Δt = + 10 ms of spike-input delay. Between these maxima, the observed change of synaptic efficacy was often bidirectional with no net change on average. Both potentiation and depression depended on calcium entry to the postsynaptic cell via L-type voltage-gated channels, presumably during the postsynaptic spike, similar to what has later been reported by Kurotani et al. (2008, cf. Figures 1E,F).

The functional implications of the observed iSTDP rule were explored in simulations of networks with dense and sparse connectivity (Haas et al., 2006). In densely connected feed-forward paths of excitatory neurons, so called synfire chains, a single interneuron was shown to successfully control runaway activity in the chain. Further, the rule scaled inhibitory strength according to the varying levels of excitatory strength and was self-stabilizing because once inhibition became strong enough, it prevented the postsynaptic spikes necessary to induce further strengthening. In a more elaborate network model with 400 excitatory and 100 inhibitory neurons and sparse, probabilistic connectivity, iSTDP in only a few synapses was sufficient to transform network-wide seizure-like activity into spatially restricted activity. When new inputs appeared at different locations, changes in synaptic strength tracked the input and grew to prevent seizure-like spread. These results show that even subtle changes in inhibitory strength can be sufficient to stabilize network dynamics.

For neocortical neurons, Vogels et al. (2011) and Luz and Shamir (2012) both hypothesized that iSTDP was the mechanism underlying the rebalancing of excitation and inhibition that occurs in A1 described above (Froemke et al., 2007). Vogels et al. (2011) studied a simplified plasticity rule, in which near-coincident pre- and postsynaptic spikes induced potentiation of the inhibitory synapse. Additionally, every presynaptic spike led to synaptic depression (Figure 1K). This learning rule was loosely based on various experimental results (Figures 1A–I) (Holmgren and Zilberter, 2001; Kilman et al., 2002; Woodin et al., 2003; Haas et al., 2006; Balena and Woodin, 2008; Hartmann et al., 2008; Woodin and Maffei, 2010), and used specifically to balance excitatory and inhibitory synaptic strengths. Vogels et al. (2011) simulated basic feed-forward networks with differentially tuned, static excitatory and untuned but plastic inhibitory synapses. Weak inhibitory synapses (those which failed to prevent postsynaptic spiking in the immediate temporal proximity of a inhibitory presynaptic spike and thus created spike pairs) were strengthened, and strong synapses (those which reliably prevented postsynaptic spikes, and thus produced only a pre-, but no postsynaptic spike) were weakened. Over time, this led to a precise, detailed balance of excitatory and inhibitory synaptic weights for each set of correlated excitatory and inhibitory signal channels as observed in Froemke et al. (2007) and others. In more general terms, the rule acts as a homeostatic mechanism in feed-forward (and also recurrent) networks that takes the inhomogeneities of the excitatory synaptic weight structure into account.

In a parallel study, Luz and Shamir (2012) have shown that the phenomenon is robust to the shape of the learning rule (Figures 1H,J,K). Asymmetric learning windows as observed by Haas et al. (2006), or even the classical asymmetric excitatory STDP (eSTDP) window (Gerstner et al., 1996; Markram et al., 1997; Bi and Poo, 1998; Song et al., 2000; Luz and Shamir, 2012) also lead to a stable balance of excitation and inhibition. The phenomenon is thus robust, as long as the learning rule obeys two fundamental requirements: Postsynaptic activity must predominantly potentiate activated inhibitory synapses, whereas in the absence of postsynaptic firing inhibitory synapses must decay. It is conceivable that the results observed by Haas et al. (2006) could also be fitted with a rule that is similar to the simplified rule proposed by Vogels et al. (2011), though with a maximum shifted by ~10 ms. Interestingly, the insensitivity of inhibitory plasticity to the exact shape of the learning window contrasts to the high sensitivity of excitatory plasticity to the formulation of the STDP curves, in particular the importance of the LTD window size relative to the LTP time window (Feldman, 2000; Song et al., 2000). This may be a consequence of the fact that inhibitory plasticity is mostly a negative feedback process, while excitatory plasticity is a form of positive feedback.

Throughout the nervous system, interneurons are frequently connected to each other through both inhibitory and gap junctional (electrical) synapses (Galarreta and Hestrin, 2001; Connors and Long, 2004). Synchrony within interneuronal networks is likely to influence the network-wide effects of their output, i.e., the many inhibitory synapses that the interneurons in these networks make on pyramidal cells. Interneuronal network synchrony has been explored for simpler networks of cells coupled by both inhibitory and electrical synapses (Chow and Kopell, 2000; Lewis and Rinzel, 2003; Kopell and Ermentrout, 2004; Pfeuty et al., 2005; Saraga et al., 2006). In many cases, inhibitory and electrical synapses play complementary roles in coordinating the activity of neurons and their inhibitory output, although electrical synapses have been shown to act as inhibitory de-synchronizers for some neurons (Vervaeke et al., 2010).

Recent findings by Haas et al. (2011) demonstrate that synchronous bursting activity, a natural form of activity for thalamic neurons that is a component of sleep spindles, in pairs of coupled interneurons depresses the electrical synapse between them. This finding implies that levels of synchrony in coupled interneuronal networks may in turn be activity-dependent. In the thalamus, for example, the reduction in electrical coupling resulting from sleep spindles would desynchronize the thalamic reticular network, and as a result the inhibition sent back to thalamus would be less temporally structured or coordinated. The effects of dynamic variations in electrical synaptic strength on the coordination of interneuronal networks have yet to be explored, and may prove to be key modulators of the impact of inhibitory plasticity across the brain.

The potential functional implications of experience-dependent inhibitory plasticity on behavior are not often easy to untangle. A recent study attempts to explore a functional consequence of inhibitory plasticity in a mouse model of communication vocalization processing (Galindo-Leon et al., 2009; Lin et al., 2013). In this context, “inhibitory plasticity” refers not necessarily to synaptic inhibition (i.e., the release of GABA), but functional inhibition in which action potential generation is suppressed in a stimulus- and state-dependent manner. Mouse pups emit ultrasonic (60–80 kHz) vocalizations that are recognized as behaviorally relevant by mothers, but not by pup-naïve virgin females (Ehret, 2005). In both animal groups, well-isolated single units recorded from auditory cortex in awake, head-restrained animals can be inhibited by, excited by, or non-responsive to a library of natural ultrasonic calls played back at superthreshold intensities. The average call-evoked excitation is not significantly different between virgins and mothers, most likely because of a wide diversity of responses produces too much variability to decipher systematic changes from one group to another. However, calls evoking inhibition show a more uniform response across different calls, making it possible to reveal significant differences in the strength of evoked inhibition between mothers and virgins. Calls elicit deeper and longer inhibition in mothers compared to virgins; and importantly, this effect is most prominent for units in core auditory cortical fields tuned to sound frequencies more lateral to those found in the ultrasonic vocalizations themselves (i.e., <50 kHz).

What are the benefits of strengthened lateral band inhibition? Galindo-Leon and colleagues explain in the framework of labeled line propagation, in which a spiking neuron conveys information to downstream targets not just by its temporal pattern of action potentials, but also by which features that neuron represents. This is thought to be true in the core auditory cortex. Cells are coarsely arranged by their preferred sound frequencies in a tonotopic spatial arrangement. Action potentials from a neuron located along this tonotopic axis thus convey that the acoustic stimulus contains some stimulus feature around the corresponding frequency. Hence, to read out the frequency content of a sound, a downstream area could assess the relative firing from each “labeled line.” In this picture, the difference (i.e., contrast) of the activity between neuronal populations becomes important for the recognition of a stimulus. However, since excitatory tuning curves at superthreshold sound levels can have large bandwidths, this population representation might be broader than expected for a narrowband signal like an ultrasonic whistle call, thereby interfering with its recognition, particularly if the call occurs in broadband background noise (Ehret, 2005). On the other hand, if neural activity from best frequency bands lateral to the call frequency were more strongly inhibited, the population contrast would be enhanced, improving the recognition of ultrasound calls.

This state-dependent change in neuronal responsiveness may be due to the forms of cellular inhibitory plasticity discussed throughout this review. However, cortical neurons have varied and distinctly mosaïic projections that complicate the simplistic picture of labeled line propagation. Though recent results seem to support the existence of a population contrast mechanism in auditory and multi-modal integration (Cohen et al., 2011), future work in awake animals is critical, and detailed modeling of the mechanisms at hand will further illuminate the functional implications.

Brain networks and circuits respond to environmental stimuli and are shaped by them. The influence of experience on the connectivity and function of sensory areas of the brain has been investigated extensively (Hubel and Wiesel, 1963, 1970; Mower et al., 1982; Merzenich et al., 1983; Fox, 1992; Hofer et al., 2009; Wittenberg, 2010). While activity-dependent excitatory plasticity is relatively well characterized as a mechanism to control the (re-)wiring of cortical circuits, inhibitory plasticity presents a much less consistent set of rules. Considering the morphological, electrical and functional diversity of interneurons in the brain, this diversity of plasticity mechanisms is not surprising and will require the use of new techniques such as genetic targeting, but also computational modeling, to be understood.

The results presented at our workshop suggest several different functional roles for inhibitory plasticity. One class of learning rules is homeostatic in nature and maintains a balance of excitation and inhibition (Haas et al., 2006; Froemke et al., 2007; House et al., 2011; Vogels et al., 2011; Luz and Shamir, 2012): neurons that receive strong excitation will also receive strong inhibition, presumably to equalize the impact of all inputs to a cell, or to reduce the differences in neuronal output between neurons receiving strong and weak excitatory drive. A different class of rules (Maffei et al., 2006; Li et al., 2008) fosters competition between neurons or synaptic inputs by increasing the inhibitory drive in response to weak (mostly subthreshold) excitation, while allowing a veto of iLTP for strong (suprathreshold) excitation. Rules of this latter type thus act as contrast enhancers, in line with the behavioral results of Galindo-Leon et al. (2009). The observed iSTDP rules of Woodin et al. (2003) and Holmgren and Zilberter (2001) (Figures 1G,I, respectively) also fit into this class, in that they reduce inhibitory drive in response to coincident pre- and postsynaptic activity. Holmgren and Zilberter (2001), however, also found a potentiation of inhibitory synapses that are activated a few hundred milliseconds after the end of a postsynaptic action potential train, introducing additional complexity that is yet to be understood. Notably, inhibitory plasticity can also change the temporal structure of neuronal responses. In particular, Gilson et al. (2012) showed that iSTDP can shape the synaptic configuration of neurons such that they become selective to specific input oscillation frequencies.

It is important to mention that plasticity of excitatory or inhibitory synapses have mostly been studied as independent phenomena. A different picture emerges in recent work that investigates the impact of inhibition on excitatory plasticity. It is becoming increasingly clear that transient neuromodulatory changes in the balance of excitation and inhibition probably form an important factor for the induction of excitatory plasticity (see, e.g., Froemke et al., 2007; Letzkus et al., 2011), potentially providing a gating mechanism that would allow to selectively learn only “behaviorally relevant” stimuli. The studies of Ormond and Woodin (2009), Wang et al. (2012), and Paille et al. (2013) show that there is a complex interaction between inhibitory and excitatory synaptic plasticity that goes beyond the idea of neuromodulation of plasticity.

As inhibitory malfunction is often implicated in neuropsychiatric diseases, a better understanding of the dynamic regulation of inhibition could also provide new insights into the pathophysiological underpinnings of diseases such as epilepsy and schizophrenia. The development of new treatments will require a careful investigation of the underlying molecular machinery, such as biophysical controllers of Cl⁻ reversal (Woodin et al., 2003; Doyon et al., 2011).

The presented results highlight a synergetic interaction between experiment and theory in the field of inhibitory plasticity. The combination of experimental characterization of the plasticity of a given synapse type and subsequent computational modeling has proven successful in evaluating the functional purposes of inhibitory plasticity and promises to be a powerful tool for the large number of future studies that will be necessary until we understand the riddles of inhibitory function and plasticity.