# US 2021/0241101 A1 · Post-training Control of the Bias of Neural Networks

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

- Number: US 2021/0241101 A1
- Document ID: US20210241101A1
- Status: Published application
- Kind: A1
- Application number: 17/140,279
- Filed: 2021-01-04
- Published: 2021-08-05
- Earliest claimed: 2020-02-04
- Assignees: DSP Group Ltd.
- Applicants: DSP Group Ltd.
- Inventors: Dimitri Lvov; Maor Shliefer; Uri Tuval
- Web page: https://patentgrep.com/patent/US20210241101A1

## CPC

- G06N 3/08: Learning methods
- G06F 18/214: Generating training patterns; Bootstrap methods, e.g. bagging or boosting
- G06F 18/241: relating to the classification model, e.g. parametric or non-parametric approaches
- G06F 18/24143: {Distances to neighbourhood prototypes, e.g. restricted Coulomb energy networks [RCEN]}
- G06F 18/2431: Multiple classes
- G06N 3/0499: Feedforward networks
- G06N 3/09: Supervised learning
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

A method for controlling a bias of a neural network, the method may include training the neural network by using a loss function that is responsive to classes of a classification process and sensitivity input values thereby setting the bias.

## Claims

### Claim 1

A method for controlling a bias of a neural network, the method comprising: training the neural network by using a loss function that is responsive to classes of a classification process and sensitivity input values thereby setting the bias.

### Claim 2 (depends on 1)

The method according to claim 1 wherein the training comprises:
(a) feeding input examples to the neural network;
(b) feeding labels indicative of one or more classes associated with each of the input examples to the neural network and to a loss function calculator;
(c) feeding the neural network and the loss function calculator with sensitivity inputs;
(d) generating neural network outputs;
(e) feeding the neural network outputs to the loss function calculator;
(f) generating outputs of the loss function calculator;
(g) feeding the outputs of the loss function calculator to a neural network configuration unit;
(h) calculating one or more weights of the neural network by the neural network configuration unit; and
(i) configuring the neural network by the neural network configuration unit.

### Claim 3 (depends on 3)

The method according to claim 3 comprising determining values of the sensitivity inputs in a random manner.

### Claim 4 (depends on 3)

The method according to claim 3 comprising re-calculating a value of a sensitivity input per each set of one or more input examples.

### Claim 5 (depends on 3)

The method according to claim 3, wherein the neural network configuration unit is an optimization unit.

### Claim 6 (depends on 1)

The method according to claim 1 wherein the loss function is a weighted variant of another loss function.

### Claim 7 (depends on 1)

The method according to claim 1 wherein the using of the loss function comprising calculating a weight that is class-dependent and sensitivity input value dependent.

### Claim 8 (depends on 7)

The method according to claim 7 wherein the weight equals a first value for a first class and equals a second value for a second class, wherein the second value is a reciprocal of the first value.

### Claim 9 (depends on 7)

The method according to claim 7 wherein the weight equals e6s for a first class and e−6s for a second class, wherein s equals u5, whereas u ranges between one and minus one.

### Claim 10 (depends on 1)

The method according to claim 1 comprising performing inference using the neural network; wherein the performing of the inference comprises providing inputs, labels and sensitivity inputs to the neural network.

### Claim 11

A non-transitory computer readable medium that stores instructions for: training a neural network by using a loss function that is responsive to classes of a classification process and sensitivity input values thereby setting the bias.

### Claim 12 (depends on 11)

The non-transitory computer readable medium according to claim 11 wherein the training comprises:
(a) feeding input examples to the neural network;
(b) feeding labels indicative of one or more classes associated with each of the input examples to the neural network and to a loss function calculator;
(c) feeding the neural network and the loss function calculator with sensitivity inputs;
(d) generating neural network outputs;
(e) feeding the neural network outputs to the loss function calculator;
(f) generating outputs of the loss function calculator;
(g) feeding the outputs of the loss function calculator to a neural network configuration unit;
(h) calculating one or more weights of the neural network by the neural network configuration unit; and
(i) configuring the neural network by the neural network configuration unit.

### Claim 13 (depends on 13)

The non-transitory computer readable medium according to claim 13 that stores instructions for determining values of the sensitivity inputs in a random manner.

### Claim 14 (depends on 13)

The non-transitory computer readable medium according to claim 13 that stores instructions for re-calculating a value of a sensitivity input per each set of one or more input examples.

### Claim 15 (depends on 13)

The non-transitory computer readable medium according to claim 13, wherein the neural network configuration unit is an optimization unit.

### Claim 16 (depends on 11)

The non-transitory computer readable medium according to claim 11 wherein the loss function is a weighted variant of another loss function.

### Claim 17 (depends on 11)

The non-transitory computer readable medium according to claim 11 wherein the using of the loss function comprising calculating a weight that is class-dependent and sensitivity input value dependent.

### Claim 18 (depends on 17)

The non-transitory computer readable medium according to claim 17 wherein the weight equals a first value for a first class and equals a second value for a second class, wherein the second value may be a reciprocal of the first value.

### Claim 19 (depends on 17)

The non-transitory computer readable medium according to claim 17 wherein the weight equals e6s for a first class and e−6s for a second class, wherein s equals u5, whereas u ranges between one and minus one.

### Claim 20 (depends on 11)

The non-transitory computer readable medium according to claim 11 that stores instructions for performing inference using the neural network; wherein the performing of the inference comprises providing inputs, labels and sensitivity inputs to the neural network.