| Activity-dependent changes in synaptic strength are widely thought to be the biological substrate of learning. Recent experimental observations indicate a previously unknown sensitivity of synaptic plasticity to the precise timing between presynaptic and postsynaptic events. This finding can extend traditional associative memory neural network models to a spatial-temporal domain. In this study, spiketiming-dependent plasticity (STDP) is analyzed in detail in a two-neuron system. The weights that evolve according to a multiplicative form of STDP converge to a stable attractor independent of the initial synaptic weights. Indeed, the asymmetrical synaptic connections asymptotically converge to values that depend on the external firing patterns, the learning kernel parameters and temporal correlation mechanisms. Thus, knowledge encoded by this intelligent system corresponds to “significant” statistical features of external firing patterns consistently, with a 99.95% probability that the relationship between inputs and a specific attractor in the weight space is not due to chance. |
