# US 2019/0378003 A1 · Neural network circuit

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2019/0378003 A1
- Document ID: US20190378003A1
- Status: Published application
- Kind: A1
- Application number: 16/550,664
- Filed: 2019-08-26
- Published: 2019-12-12
- Earliest claimed: 2017-03-03
- Assignees: Denso Corporation
- Applicants: Denso Corporation
- Inventors: Shigeki OTSUKA; Irina KATAEVA
- Web page: https://patentgrep.com/patent/US20190378003A1

## CPC

- G06N 3/063: using electronic means
- G06F 11/2028: {eliminating a faulty processor or activating a spare}
- G11C 11/54: using elements simulating biological cells, e.g. neuron
- G11C 13/00: Digital stores characterised by the use of storage elements not covered by groups G11C11/00, G11C23/00, or G11C25/00
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G11C: STATIC STORES (semiconductor memory devices H10B)

## Abstract

A neural network circuit is provided. The neural network circuit includes a memory device including memristors connected in a matrix, a controller arranged to control a voltage application device to perform writing, deleting and reading data in the memory device, multiple current-to-voltage (I-V) conversion amplifier circuits arranged to convert currents flowing through the memory elements into voltages and outputting the voltages, and multiple current adjusters respectively corresponding to the I-V conversion amplification circuits, each current adjuster being arranged to adjust a total current value input to a corresponding I-/V conversion amplification circuit to zero.

## Claims

### Claim 1

A neural network circuit comprising:
a memory device including memristors being variable resistance elements connected in a matrix, wherein the memristors act as memory elements and resistance values of the memristors provide data of the memory device;
a voltage application device arranged to apply bias voltages to the memory device;
a controller arranged to control the voltage application device to perform writing, deleting and reading the data in the memory device, wherein the writing reduces the resistance value of one or more of the memory elements and the deleting increases the resistance value of one or more of memory elements;
a plurality of current-to-voltage (I-V) conversion amplifier circuits arranged to convert currents flowing through the memory elements into voltages and outputting the voltages; and
a plurality of current adjusters respectively corresponding to the plurality of I-V conversion amplification circuits, each current adjuster being arranged to adjust a total current value input to a corresponding I-N conversion amplification circuit to zero.

### Claim 2 (depends on 1)

The neural network circuit according to claim 1, wherein
the current adjuster includes a current source arranged at an input part of the corresponding I-V conversion amplification circuit.

### Claim 3 (depends on 2)

The neural network circuit according to claim 2, wherein
a temperature characteristic of a reference current flowing inside the current source is set equal to a temperature characteristic of an input current of the I-V conversion amplification circuit.

### Claim 4 (depends on 1)

The neural network circuit according to claim 1, wherein
the current adjuster includes a series circuit of a resistance element and a voltage source, the series circuit being arranged at an input part of the corresponding I-V conversion amplification circuit.

### Claim 5 (depends on 1)

The neural network circuit according to claim 1, further comprising
a resistance value adjuster arranged to reset a resistance value of a corresponding memory element in cases where: after adjustment is performed with the current adjuster, it is detected that the total current value is not zero.

### Claim 6 (depends on 1)

The neural network circuit according to claim 1, further comprising
a differential operation device arranged to provide a difference between outputs of any two of the current adjusters.

### Claim 7 (depends on 1)

The neural network circuit according to claim 1, wherein
the memory device has a differential configuration,
the neural network circuit further comprising
a differential operation device arranged to provide a difference between outputs of two of the I-V conversion amplifier circuits constituting a differential pair.

### Claim 8 (depends on 7)

The neural network circuit according to claim 7, further comprising
a resistance value adjuster arranged to, if the difference is not zero, adjust the resistance value of the memory elements corresponding to the differential pair to make the difference zero.

### Claim 9 (depends on 8)

The neural network circuit according to claim 8, further comprising
an absolute value adjuster arranged to make adjustment to make an absolute value of the difference zero when it is unable to make the adjustment using the resistance value adjuster.

### Claim 10 (depends on 9)

The neural network circuit according to claim 9, wherein:
the differential operation device includes a positive side input terminal and a negative side input terminal and is arranged to enable the positive side input terminal to act as the negative side input terminal and the negative side input terminal to act as the positive side input terminal;
the absolute value adjuster performs rewriting so as to interchange a resistance value of a memory element connected to the positive side input terminal and a resistance value of a memory element connected to the negative side input terminal; and
the absolute value adjustment makes the absolute value of the difference zero through causing the positive side input terminal and the negative side input terminal, respectively, to act as the negative side input terminal and the positive side input terminal.

### Claim 11 (depends on 7)

The neural network circuit according to claim 7, further comprising:
spare memory elements to be replaced with differential memory elements including an abnormality-detected memory element, wherein when the difference between the outputs is not zero, the resistances of the differential memory elements are adjusted to be equal to each other; and
an abnormality handling device arranged to replace the differential memory elements including the abnormality-detected memory element with the spare memory elements through setting the resistance values of the differential memory elements connected to the same input terminal to a maximum value except for the abnormality-detected memory element.

### Claim 12 (depends on 11)

The neural network circuit according to claim 11, further comprising:
a resistance value adjuster that, when the difference between the outputs is not zero, adjusts the resistance value of the abnormality-detected memory element so as to make the difference zero,
when it is unable to make the adjustment with the resistance value adjuster, the abnormality handling device replaces the differential memory elements including the abnormality-detected memory element with the spare memory elements.
