Back to results
US 12,135,497 B2Granted
Random weight initialization of non-volatile memory array
AssigneeInternational Business Machines Corporation
InventorsCheng Chi, Takashi Ando, Reinaldo Vega, Praneet Adusumilli
Life of this patent
Estimating the term…
Abstract
A memory device is provided. The memory device includes a main feature disposed beneath a surface of a photolithographic mask. The memory device further includes at least one Sub-Resolution Assistant Feature (SRAF) proximate to the main feature beneath the surface. The main feature has an electrical conductivity based on an area relationship with the at least one SRAF.
Claims20
A device, comprising:
- a main feature disposed on a surface of a photolithographic mask;
- and
- Sub-Resolution Assistant Features (SRAFs) proximate to the main feature on the surface, the main feature being sized to generate a device area on a memory device based on an area dependence relationship with the SRAFs, the SRAFs to introduce programmable noise that provide a distribution of weight initialization for the main feature to provide initialization weights for a neural network.
A method for forming a device, comprising:
- forming a main feature disposed on a surface of a photolithographic mask;
- and
- forming Sub-Resolution Assistant Features (SRAFs) proximate to the main feature on the surface, the main feature being sized to generate a device area on a memory device based on an area dependence relationship with the SRAFs, the SRAFs to introduce programmable noise that provide a distribution of weight initialization for the main feature to provide initialization weights for a neural network.
A device, comprising:
- a plurality of main features disposed on a surface of a photolithographic mask;
- and
- respective Sub-Resolution Assistant Features (SRAFs) proximate to each of the plurality of main features on the surface, each of the plurality of main features being sized to generate a device area on a memory device based on an area dependence relationship with the SRAFs proximate thereto, the SRAFs to introduce programmable noise that provide a distribution of weight initialization for the plurality of main features to provide initialization weights for a neural network.
Dependencies are inferred from the claim text.