STDP Installs in Winner-Take-All Circuits an Online Approximation to Hidden Markov Model Learning
(A) The structure of the network. It consists of excitatory neurons (blue) that receive feedforward inputs (green synapses) and lateral excitatory all-to-all connections (blue synapses). Interneurons (red) install soft winner-take-all behavior by injecting a global inhibition to all neurons of the circuit in response to the network's spiking activity. (B) The Bayesian network representing the HMM over time steps. The prediction model (blue arrows) is implemented by the lateral synapses. It determines the evolution of the hidden states over time. The observation model (green arrows) is implemented by feedforward connections. The inference task for the HMM is to determine a sequence of hidden states (white), given the afferent activity (gray). (C) The STDP window that is used to update the excitatory synapses. The synaptic weight change is plotted against the time difference between pre- and postsynaptic spike events.