Silicon Valley based - Neural Systems Corporation (NSC) has developed a technology called trainable digital logic (TDL), which performs many of the functions performed by artificial neural networks. TDL circuitry typically is significantly simpler and orders of magnitude faster than the equivalent neural network. General purpose TDL pattern recognizers can be constructed using FPGAs or as part of a chip that performs additional functions. Development funding was provided by the Advanced Technology Program of the National Institute of Standards and Technology.
Almost all conventional neural networks are implemented on a digital computer where the computer input is a digital word (or words) and the output is binary (or a binary word). Thus, the computer implements what is generally known as a 'switching function'. It is known that all switching functions can be implemented with binary logic. TDL, which uses binary logic, is trained to implement the complicated switching functions that are equivalent to those generated by a neural network computer. During training NSC algorithms specify the logic and architecture in a compact form. Additional FPGA specific algorithms can be used to implement the logic.
Conventional neural network algorithms implement large sets of interconnected neuron analogs. Each analog requires much arithmetic, and solutions of non-linear functions. The result is complexity and reduced speed. The TDL, in contrast, uses no arithmetic and no non-linear functions. Typically, it is orders of magnitude faster and significantly less complex. For non-neural network-like applications FPGA based TDL can be used in a similar manner to implement a complex switching function (or switching functions) by training in real time on sets of binary words. Using this approach, the TDL pattern recognizer hardware is capable of making millions of decisions a second.
NSC is interested in aiding in the development of products that may contain TDL. For contact information and additional details about the technology visit the NSC Web site at: