P290 Rate-based versus spike-triggered contributions in spike-timing–dependent synaptic plasticity
Jakob Stubenrauch*1,2, Benjamin Lindner1,2
1Bernstein Center for Computational Neuroscience Berlin, Philippstraße 13, Haus 2, 10115 Berlin, Germany 2Physics Department of Humboldt University Berlin, Newtonstraße 15, 12489 Berlin, Germany
*Email: jakob.stubenrauch@rwth-aachen.de IntroductionSpike-timing-dependent plasticity (STDP) has long been proposed as a phenomenological model class for synaptic learning [1], yet most theoretical frameworks of learning reduce plasticity to effectively rate-based descriptions. The short window of spike-pairs contributing to STDP of around 20ms [1] however points to the relevance of precise post-synaptic spike-responses. We investigate this timing-sensitive aspect of plasticity by dissecting synaptic dynamics into two contributions: spike-pairs that fall into the STDP window by rate-dependent coincidence versus those occurring through direct causation—a crucial distinction that reflects fundamentally different learning mechanisms.
MethodsWe develop a theoretical framework for the drift and diffusion of synaptic weights under STDP. We leverage established results [2,3] on the response of leaky integrate-and-fire (LIF) neurons, mean field theory of spiking networks [4], and recent advances in shot-noise theory [5]. Specifically, we derive a Langevin equation that describes the stochastic evolution of synaptic weights. This framework naturally subdivides the synaptic dynamics into rate-based and correlated contributions. The theory is applied to synapses that deliver Poissonian spikes into a recurrent network of LIF neurons, for which it captures per-realization the population mean and variance of the weights. The theory is tested against simulations of spiking neurons. ResultsOur analysis quantifies and dissects the dynamics of synaptic weights. The contribution from correlated response—neglected in effectively rate-based descriptions—increases with the mean synaptic weight and becomes significant even at modest weights where ~20 concurrent input spikes are needed to reliably elicit action potentials. We apply the theory to characterize a supervised training paradigm mimicking memory consolidation. In this paradigm, the drift and diffusion derived by the theory capture the encoding strength and decay of memory traces and, more importantly, manage to attribute these to rate-based and correlation-dependent contributions, respectively.
DiscussionThe precise response of spiking neurons matters for plasticity if synaptic weights are large enough. As we demonstrate, this effect can have a large impact on the success or failure of associative learning. Based on our work, it is thus possible to judge under which circumstances STDP’s strong tuning to closely succeeding spikes is important. Correspondingly, when capturing the rate-based effects of STDP, one may overlook crucial aspects of learning. Future research should extend this approach to different neuron models, network architectures, and training paradigms. Results should be tested experimentally. Last, it would be of high interest to extend the framework to multiple populations and to recurrent plasticity.
Acknowledgements This work has been funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation), SFB1315 B01, Project-ID No. 327654276 to B. L. References 1.https://doi.org/10.1523/jneurosci.18-24-10464.1998 2.https://doi.org/10.1103/PhysRevLett.86.2186 3.https://doi.org/10.1103/PhysRevLett.86.2934 4.https://doi.org/10.1023/A:1008925309027 5.https://doi.org/10.1103/PhysRevX.14.041047
