I’ll add this one in as well: Hopefully it’s useful.
A controller is a controller of an array including a neuromorphic element that multiplies a weight based on a value of a variable characteristic by a signal, and includes a control unit that controls the characteristic of the neuromorphic element by using a discretization step size obtained so...
patents.justia.com
CONTROLLER OF ARRAY INCLUDING NEUROMORPHIC ELEMENT, METHOD OF ARITHMETICALLY OPERATING DISCRETIZATION STEP SIZE, AND PROGRAM
Feb 19, 2018 -
TDK CORPORATION
A controller is a controller of an array including a neuromorphic element that multiplies a weight based on a value of a variable characteristic by a signal, and includes a control unit that controls the characteristic of the neuromorphic element by using a discretization step size obtained so that a predetermined condition for reducing an error or a predetermined condition for improving accuracy is satisfied on the basis of a case where a true value of the weight obtained with a higher accuracy than a resolution of the characteristic of the neuromorphic element is used and a case where a discretization step size which is set for the characteristic of the neuromorphic element is used.
This is a method of correcting errors in analog neurons - again a pre-baked NPU.
[0006] As an example, Patent Literature 1 discloses a method of loading a weight (connection weight) obtained through real number value simulation into a circuit chip of a neural network including a discrete value synapse device in a spike-type neural network, and the circuit chip includes a neuromorphic element (see Patent Literature 1).
[0007] However, one problem occurring in a case where a neuromorphic element is applied to a neural network is a
resolution of a resistance change. That is, a resistance of a neuromorphic element is not changed in a completely analog manner, but has discrete values like in a quantization step, and thus the use of a neuromorphic element in a neural network may result in the occurrence of a quantization error and deterioration of performance such as in identification.
[0008] For example, in a case where a neuromorphic element is used for a weight storage function and a weight updating function in a neural network, expressiveness of variables is insufficient as compared with a real-number-based simulation using a computer, and thus deterioration of identification performance or an increase in a period of time required until weight update function reaches convergence.
19 . An arithmetic operation method of arithmetically operating a discretization step size of a characteristic of a neuromorphic element for an array including the neuromorphic element that multiplies a weight based on a value of a variable characteristic by a signal, the arithmetic operation method comprising:
a step of arithmetically operating a true value of the weight with higher accuracy than a resolution of the characteristic of the neuromorphic element; and
a step of arithmetically operating a discretization step size so that a predetermined condition for reducing an error or a predetermined condition for improving accuracy is satisfied on the basis of a case where the true value of the weight is used and a case where the discretization step size which is set for the characteristic of the neuromorphic element is used.