Hebbian learning has become a very influential paradigm, which states that learning in neural networks is achieved by modifying synaptic connections between neurons in proportion to their correlated activity (Gerstner and Kistler 2002). While this process is desirable to detect relevant inputs to be learned, it also represents a positive feedback mechanism that becomes discriminatory when different synapses do not have equal starting points to begin with. For example, activation of a “strong” synapse from neuron A onto neuron B will increase the activity of neuron B, resulting in an increase in the correlation between the two neurons and hence an increase in its own strength—and conversely, a further weakening of weak synapses. Starting points are unequal when applying Hebbian learning to synapses located on dendrites, as are most excitatory synapses in the CNS. Postsynaptic potentials are attenuated as they travel from the synapse toward the soma in a manner that depends on the cable structure and the membrane properties of dendrites. Thus distal synapses are weaker to start with, compared to more proximal synapses, in their ability to influence neuronal output via the axon. Facing such a situation, distal dendritic synapses might as well give up completely. Not only are they attenuated, but according to Hebbian learning their destiny is extinction.
Au contraire. In several cell types in which synaptic conductance has been measured directly as a function of the distance of the synapses from the soma, it was at least constant on average (Williams and Stuart 2002) or even increased with distance from the soma (Magee and Cook 2000; for review, see Häusser 2001). This distance-dependent scaling of the synaptic conductance appears to counterbalance the voltage attenuation, rendering excitatory postsynaptic potential (EPSP) amplitude at the soma insensitive to its dendritic origin, a situation that has been called “dendritic democracy” (Häusser 2001; but see London and Segev 2001). But how can a synapse possibly tell its distance from the soma to scale its conductance properly?
In this issue of the Journal of Neurophysiology (p. 2273–2280), Rumsey and Abbott provide an elegant demonstration of how synapses can solve this problem and rebuild dendritic democracy using an anti-Hebbian spike-timing-dependent plasticity (anti-STDP) rule. The key to their method relies on the important distinction between “synaptic strength” and “synaptic efficacy.” The first reflects the biophysical parameters of the synapse that are modified by the learning rule (e.g., postsynaptic conductance, presynaptic release probability), whereas the second measures the effect of the synapse on neuronal output. When a presynaptic spike precedes the postsynaptic spike, the synaptic strength is reduced according to the anti-STDP rule and vice versa. This makes the algorithm sensitive to the implicit synaptic efficacy: a proximal dendritic synapse that initially tends to increase the probability of the postsynaptic neuron to fire will be weakened because its presynaptic spike will tend to occur before the postsynaptic spikes, whereas a distal synapse will become stronger. When applied to dendritic neurons, anti-STDP therefore results in distance-dependent synaptic scaling, leading to equalization of synaptic efficacies (rather than equalization of somatic EPSP amplitudes).
While this is a simple and efficient solution to a potentially serious problem, Rumsey and Abbott point out that many questions remain. For example, how is anti-STDP implemented as a biophysical mechanism? How can anti-STDP coexist peacefully with (Hebbian) STDP mechanisms that strengthen correlated synaptic inputs and would therefore be expected to undo the work of anti-STDP? Probably a range of Hebbian and homeostatic mechanisms (Turrigiano and Nelson 2000) is operating in parallel, at different time scales, and synapses between different types of neurons may be governed by different STDP mechanisms. This may apply even to synapses at different dendritic locations on the same postsynaptic neuron because the long-term depression (LTD) part of the anti-STDP timing rule depends on retrograde signaling via the backpropagating action potential (bAP), which informs the dendritic synapses that the axon has fired (Stuart et al., 1997). However, backpropagation is decremental in most cell types (Stuart et al. 1997; Vetter et al. 2001; Waters et al. 2003), and the decreasing amplitude of the bAP is likely to influence the biophysical mechanisms underlying the timing rule (Sourdet and Debanne 1999). In fact, the amplitude and waveform of the bAP might be an alternative source of information from which synapses can read out their dendritic position. It will not be easy to validate experimentally that anti-STDP is responsible for distance-dependent synaptic scaling because it is difficult to measure synaptic efficacy and because the synaptic changes are probably happening slowly.
In the neocortex, synaptic connections between pairs of neurons are typically made by several individual synaptic contacts (Markram et al., 1997a). Because they share the same presynaptic axon, their activity is highly correlated, and it is their joint efficacy that would be equalized to that of other presynaptic axons in the model of Rumsey and Abbott. This kind of equalization is not observed experimentally, however, as the unitary somatic EPSP amplitude of synaptic connections between the same types of neurons varies over more than an order of magnitude (Markram et al., 1997a). This could be due to (Hebbian) STDP operating in parallel (Markram et al., 1997b), and of course due to different patterns of correlations between different presynaptic axons. Intriguingly, different synaptic contacts within a given connection between a pair of neurons, which share similar histories of pre- and postsynaptic activity, do show a normalization of synaptic strength (Koester et al., 2003). The work of Rumsey and Abbott should provide a stimulus for further experimental and theoretical research into STDP timing rules and location dependence of synaptic plasticity.
- Copyright © 2004 by the American Physiological Society