Method and apparatus for training memristive learning systems
Disclosed is a method for training memristive learning systems (MLSs) using stochastic learning algorithms and the training system apparatus designed to implement the stochastic learning algorithms.
1. A method for training a memristive learning system comprising:
identifying a task that the memristive learning system is to be trained to perform;
identifying a cost function;
choosing a deterministic update equation;
deriving a stochastic update equation from the deterministic update equation by converting the deterministic update equation into a stochastic bit stream;
designing a training system hardware comprising reduced hardware training circuitry, cost and power consumption required to perform the task to implement the derived stochastic update equation by applying stochastic write voltages to the memristors where the stochastic write voltages are generated from the stochastic bit stream;
coupling the memristive learning system to the training system; and
training the memristive learning system to provide a trained memristive learning system while minimizing the hardware training circuitry, cost and power consumption required to train the memristive learning system.
2. The method of claim 1 , wherein the task comprises image classification, time series prediction, function approximation, coincidence detection, or clustering.
3. The method of claim 1 , wherein the deterministic equation is supervised, unsupervised, or semi supervised.
4. The method of claim 1 , wherein the cost function comprises cross entropy or mean square error cost function.
5. The method of claim 1 , wherein deriving a stochastic update equation from the deterministic update equation comprises changing each continuous analog value to a digital value drawn from a probability distribution.
6. The method of claim 1 , wherein training the memristive learning system comprises applying voltages to at least one memristor of the memristive learning system and the adjusting the conductances of the at least one memristor based on the training system implementation of the stochastic equation in at least one of a supervised, unsupervised, and semi supervised fashion in order to change the learning system's functionality.