A unified model of short- and long-term plasticity: Effects on network connectivity and information capacity
Activity-dependent synaptic plasticity is a fundamental learning mechanism that shapes the connectivity and activity of neural circuits. Existing computational models of Spike-Timing-Dependent Plasticity (STDP) capture long-term synaptic changes with varying degrees of biological detail. A common approach is to neglect the influence of short-term dynamics on long-term plasticity, which may be an oversimplification for certain neuron types. Thus, there is a need for new models to investigate how short-term dynamics influence long-term plasticity. To address this gap, we introduce a novel phenomenological model, the Short-Long-Term STDP (SL-STDP) rule, which directly integrates the Tsodyks-Markram model of short-term dynamics with postsynaptic long-term plasticity. We fit the new model to recordings from layer 5 of the visual cortex and study how short-term plasticity affects the firing rate frequency dependence of long-term plasticity in a single synapse. Our analysis revealed that the pre- and postsynaptic frequency dependence of long-term plasticity plays a crucial role in shaping the self-organization of recurrent neural networks (RNNs) and their information processing through the emergence of sinks and source nodes. We applied the SL-STDP rule to RNNs and found that neurons in the SL-STDP network self-organize into distinct firing rate clusters, stabilizing the dynamics. We extended the experiments by including homeostatic balancing, namely weight normalization and excitatory-to-inhibitory plasticity, and observed differences in degree correlations between the SL-STDP network and a network without direct coupling between short-term and long-term plasticity. Finally, we evaluated how the modified connectivity affects the networks information capacity in reservoir computing tasks. The SL-STDP rule outperformed the uncoupled system in the majority of tasks, and including excitatory-to-inhibitory facilitating synapses further improved information capacity. Our study demonstrates that short-term dynamics-induced changes in the frequency dependence of long-term plasticity play a pivotal role in shaping network dynamics and link synaptic mechanisms to information processing in RNNs. Author summaryThe brain is a complex organ capable of developing, adapting and learning throughout life. Learning and development of the brain is facilitated by several different plasticity mechanisms that act on different brain areas and timescales. Computational modeling of these plasticity mechanisms not only help us to understand the working principles of the brain but can provide us useful algorithms for future computing devices. In this work, we develop a new activity-dependent plasticity model that acts locally in a synapse and combines two timescales of plasticity into one set of equations. We investigate the effects of this new synapse model on recurrently connected neural networks and observe changes in neural activity and connectivity compared to traditional approaches. We further evaluate the model by performing information capacity tests that are related to the working memory of the circuit. We find improved information capacity, indicating enhanced computational performance of the proposed model. Our study emphasizes how combining the two timescales of plasticity may be important for the development of neural circuits and working memory.