Exploration of learning in an associative magnetic processor

The authors previously (1987) proposed a physical implementation of an amorphous magnetic processor functioning as a fast associative memory with massive information storage. An amorphous magnetic material, such as a spin glass film, is used. Magnetic particles, which orient along or against the direction of local magnetic fields (called spins) in a film, play the role of neurons or units in a neural net. Magnetic interactions in the film serve the role of weights in a connectionist neural net model. The film is taught a series of pattern vectors imposed by local laser heating in a magnetic field by allowing the magnetic spins of the system to move and seek a self-consistent energy minimum. The energy landscape created effectively stores a set of pattern vectors and permits their subsequent retrieval. The authors explore two training procedures (learning rules) and discuss computer simulations by which the effectiveness of learning in this machine can be studied. In addition to its use as associative memory, its use for data compression and unsupervised feature direction is also discussed.<>