1: \begin{abstract}

2: We propose a particularly structured Boltzmann machine, which we refer

3: to as a dynamic Boltzmann machine (DyBM), as a stochastic model of a

4: multi-dimensional time-series.  The DyBM can have infinitely many

5: layers of units but allows exact and efficient inference and learning

6: when its parameters have a proposed structure.  This proposed

7: structure is motivated by postulates and observations, from biological

8: neural networks, that the synaptic weight is strengthened or weakened,

9: depending on the timing of spikes (i.e., spike-timing dependent

10: plasticity or STDP).  We show that the learning rule of updating the

11: parameters of the DyBM in the direction of maximizing the

12: likelihood of given time-series can be interpreted as STDP with

13: long term potentiation and long term depression.  The learning rule

14: has a guarantee of convergence and can be performed in a distributed

15: matter (i.e., local in space) with limited memory (i.e., local in

16: time).

17: \end{abstract}

18: