A hardware-implementable algorithm for learning nonlinear functions

An algorithm is presented which allows continuous functions to be learned by a neural network using spike-based stochastic reinforcement training. The algorithm may be implemented in hardware by probabilistic random access memory (pRAM) nodes. The addition of output transformation modules which implement a squashing function (with trainable 'inverse temperature' parameter /spl beta/) allows pRAM nets to act as universal approximators; the presence of higher-order terms in the pRAM output function may lead to particularly compact solutions to difficult problems in nonlinear function learning.