Associative Memory

Uses perdeterminded patterns of sign flipping followed by the fast Walsh Hadamard transform to create Random Projections (RPs.)

The RPs are then binarized to give a Locality Sensitive Hash.

The bits of the hash are regarded as +1 or -1, weighted and summed to give a recalled value.

This is done in as many dimensions as wanted. To train, a recall is done and the error calculated.

The error is divided by the number of bits used and that is added or subtracted from each weight as appropriate to make the error zero.

The effect on previously stored memories to to contaminate them with some Gaussian noise.

By repeated training on the examples the Gaussian noise can be driven to zero if the capacity of the AM is big enough.

WHT 1 WHT 2 WHT 3 WHT 4 WHT 5 Synopsis

Auto-Associative Example.

Mouse click to select examples, press 1 to train, 0 to clear.

Free Web Hosting