Local Learning in Probabilistic Networks with Hidden Variables

Probabil istic networks which provide compact descriptions of complex stochastic relationships among several random variables are rapidly becoming the tool of choice for uncertain reasoning in artif icial intelligence We show that networks wi th fixed structure containing hidden variables can be learned automatically f rom data using a gradient-descent mechanism similar to that used in neural networks We al io extend the method to networks wi th intensionally represented d is t r i butions, inc luding networks wi th continuous variables and dynamic probabil ist ic networks Because probabil ist ic networks provide expl ici t representations of causal structure human experts can easily contribute pnor knowledge to the training process, thereby signif icantly improving the learning rate Adaptive probabilistic networks (APNs) may soon compete directly w i th neural networks as models in computational neuroscience as wel l as in industrial and financial applications