Learning transformed prototypes (LTP) - A statistical pattern classification technique of neural networks

Abstract

A statistical pattern recognition algorithm called learning transformed prototypes (LTP) is developed for probabilistic RAM (pRAM) neural networks, With LTP the pRAM net learns to map statistically the input sets to the output prototypes, or codebook vectors, in the binary domain. The method allows the pRAM net to self-organise the codebook vectors in the output space of arbitrary dimension. The similarities and differences of LTP with those algorithms such as LVQ (learning vector quantisation), SOFM (self-organised feature maps) and pRAM reinforcement learning are discussed. The training data processed in the method is the input-output spike series of the neural net, therefore the technique can be built into a hardware system with the currently available pRAM learning chips.