Content deleted Content added
No edit summary |
|||
Line 21:
In [[computer science]], '''zero instruction set computer''' ('''ZISC''') refers to a [[computer architecture]] based solely on [[pattern matching]] and absence of [[instruction (computer science)|(micro-)instructions]] in the classical{{huh|reason=What would be an example of a non-classical micro-instruction?|date=December 2016}}<!-- some confusion here could probably be resolved by updating the instruction article --> sense. These chips are known for being thought of as comparable to the [[neural network]]s, being marketed for the number of "synapses" and "neurons".<ref name="BrainChip"/> The [[acronym and initialism|acronym]] ZISC alludes to [[reduced instruction set computer]] (RISC).{{fact|date=December 2016}}
ZISC is a hardware implementation of [[Kohonen network]]s (artificial neural networks) allowing massively parallel processing of very simple data (0 or 1). This hardware implementation was invented by Guy Paillet,{{Citation needed|date=December 2018}} and Pascal Tannhof (IBM)<ref>https://www.researchgate.net/profile/Pascal-Tannhof</ref>, developed in cooperation with the IBM chip factory of [[Essonnes]], in France, and was commercialized by IBM.
The ZISC architecture alleviates the [[memory bottleneck]]{{clarify|date=December 2016}} by blending pattern memory with pattern learning and recognition logic.{{how|date=December 2016}} Their massively parallel computing solves the {{Clarify|text="[[Winner-take-all in action selection|winner takes all problem in action selection]]"|post-text=from [[Winner-take-all (computing)|Winner-takes-all]] problem in [[Artificial neural network|Neural Network]]s|reason=Per [https://web.archive.org/web/20170101001452/https://pdfs.semanticscholar.org/1e0c/54bd88223e009997a04dcd2a0f3fa0af3848.pdf source], [[Winner-take-all (computing)|Winner-takes-all]] is defined as a different principle from [[Winner-take-all in action selection]], but both are relevant to [[Artificial neural network|Neural Network]]s|date=December 2016}} by allotting each "neuron" its own memory and allowing simultaneous problem-solving the results of which are settled up disputing with each other.<ref name="Gigaom"/>
| |||