# US 2021/0279591 A1 · Minimum Deep Learning with Gating Multiplier

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

- Number: US 2021/0279591 A1
- Document ID: US20210279591A1
- Status: Published application
- Kind: A1
- Application number: 16/809,096
- Filed: 2020-03-04
- Published: 2021-09-09
- Earliest claimed: 2020-03-04
- Assignees: Google Inc.
- Applicants: Google LLC
- Inventors: Gil Shamir
- Web page: https://patentgrep.com/patent/US20210279591A1

## CPC

- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 3/04: Architecture, e.g. interconnection topology
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G06N 3/048: Activation functions
- G06N 3/0495: Quantised networks; Sparse networks; Compressed networks
- G06N 3/0499: Feedforward networks
- G06N 3/082: modifying the architecture, e.g. adding, deleting or silencing nodes or connections
- G06N 3/09: Supervised learning
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

Systems and methods according to the present disclosure can employ a computer-implemented method for inference using a machine-learned model. The method can be implemented by a computing system having one or more computing devices. The method can include obtaining data descriptive of a neural network including one or more network units and one or more gating paths, wherein each of the gating path(s) includes one or more gating units. The method can include obtaining data descriptive of one or more input features. The method can include determining one or more network unit outputs from the network unit(s) based at least in part on the input feature(s). The method can include determining one or more gating values from the gating path(s). The method can include determining one or more gated network unit outputs based at least in part on a combination of the network unit output(s) and the gating value(s).

## Claims

### Claim 1

A computing system for performing gating-based regularization of a neural network, the computing system comprising:
one or more processors; and
one or more non-transitory computer-readable media that collectively store:
a neural network, the neural network comprising:
a gated network unit, the gated network unit comprising one or more network parameters; and
a gating path associated with the gated network unit, wherein the gating path comprises one or more gating units, wherein each of the one or more gating units comprises one or more gating parameters, wherein the gating path is configured to produce a gating value;
wherein a gated output of the gated network unit comprises an intermediate output of the gated network unit multiplied by the gating value; and
instructions that, when executed by the one or more processors, cause the computing system to perform operations to train the neural network based on one or more training examples, wherein the operations comprise, for each of the one or more training examples:
determining a gradient of a loss function with respect to at least one of the one or more network parameters and one or more gating parameters; and
updating a respective value of at least one of the one or more network parameters and the one or more gating parameters based on the gradient of the loss function.

### Claim 2 (depends on 1)

The computing system of claim 1, wherein the one or more gating units comprises at least one of a benefit path configured to produce a benefit score associated with the gated network unit, a scaling function configured to produce the gating value based at least in part on the benefit score, or a clipping function configured to clip the gating value based at least in part on a clipping threshold.

### Claim 3 (depends on 2)

The computing system of claim 2, wherein the benefit path comprises at least one of a benefit bias unit or a benefit link weight, wherein the one or more gating parameters comprises a benefit score bias, wherein the benefit bias unit is configured to store the benefit score bias, and wherein the benefit path produces the benefit score based at least in part on the benefit score bias, wherein the benefit path comprises at least one of a stateful benefit bias unit or a stateful benefit link weight.

### Claim 4 (depends on 2)

The computing system of claim 2, wherein the neural network comprises one or more network layers, each of the one or more network layers comprising one or more network units, wherein a first network layer of the one or more network layers comprises the gated network unit, and wherein the benefit path comprises a weighted sum of inputs, wherein the weighted sum of inputs comprises outputs of the one or more network units.

### Claim 5 (depends on 2)

The computing system of claim 2, wherein the benefit path comprises one or more benefit path layers, each of the benefit path layers comprising one or more benefit units.

### Claim 6 (depends on 5)

The computing system of claim 5, wherein at least one of the one or more benefit path layers comprises a bottleneck layer, wherein a dimensionality of the bottleneck layer is less than a dimensionality of a preceding layer of the one or more benefit path layers.

### Claim 7 (depends on 1)

The computing system of claim 1, wherein the neural network comprises a second gated network unit, wherein a second gated output of the second gated network unit comprises a second intermediate output of the second gated network unit multiplied by the gating value.

### Claim 8 (depends on 1)

The computing system of claim 1, wherein at least one of the one or more gated network units and one or more gating units are stateful gating units that are respective to a state of the at least one gated network unit, wherein the at least one gated network unit having the state comprises at least one of an embedding unit, a link weight, or a bias unit.

### Claim 9 (depends on 1)

The computing system of claim 1, wherein the one or more network units comprises an embedding vector, and wherein the gating value is associated with each embedding component in the embedding vector.

### Claim 10 (depends on 2)

The computing system of claim 2, wherein the scaling function comprises one of a sigmoid function, a half-sigmoid function, or shifted piecewise smooth activation function.

### Claim 11 (depends on 2)

The computing system of claim 2, wherein the scaling function comprises a self-gating scaling function.

### Claim 12 (depends on 1)

The computing system of claim 1, wherein the one or more gating parameters are learned during training of the neural network.

### Claim 13 (depends on 1)

The computing system of claim 1, wherein the one or more gating units are employed as an activation function for the gated network unit.

### Claim 14

A computer-implemented method for performing inference using a machine-learned model, the computer-implemented method comprising:
obtaining, by a computing system comprising one or more computing devices, data descriptive of a neural network comprising:
one or more network units; and
one or more gating paths, each of the one or more gating paths associated with each of the one or more network units, wherein each of the one or more gating paths comprises one or more gating units;
obtaining, by the computing system, data descriptive of one or more input features;
determining, by the computing system, one or more network unit outputs from the one or more network units based at least in part on the one or more input features;
determining, by the computing system, one or more gating values from the one or more gating paths; and
determining one or more gated network unit outputs based at least in part on a combination of the one or more network unit outputs and the one or more gating values.

### Claim 15 (depends on 14)

The computer-implemented method of claim 14, wherein the one or more gating units comprises at least one of a benefit path configured to produce a benefit score associated with the gated network unit, a scaling function configured to produce the gating value based at least in part on the benefit score, or a clipping function configured to clip the gating value based at least in part on a clipping threshold.

### Claim 16 (depends on 15)

The computer-implemented method of claim 15, wherein the benefit path comprises at least one of a benefit bias unit or a benefit link weight, wherein the one or more gating parameters comprises a benefit score bias, wherein the benefit bias unit is configured to store the benefit score bias, and wherein the benefit path produces the benefit score based at least in part on the benefit score bias, wherein the benefit path comprises at least one of a stateful benefit bias unit or a stateful benefit link weight.

### Claim 17 (depends on 15)

The computer-implemented method of claim 15, wherein the neural network comprises one or more network layers, each of the one or more network layers comprising one or more network units, wherein a first network layer of the one or more network layers comprises the gated network unit, and wherein the benefit path comprises a weighted sum of inputs, wherein the weighted sum of inputs comprises outputs of the one or more network units.

### Claim 18

A computer-implemented method of deploying a machine-learned model, the computer-implemented method comprising:
obtaining, by one or more computing devices, data descriptive of a neural network comprising one or more network units and one or more gating units, the one or more gating units comprising one or more gating paths associated with the one or more network units;
training, by the one or more computing devices, the neural network to learn one or more network parameters of the one or more network units and one or more gating parameters of the one or more gating units;
sparsifying, by the one or more computing devices, the neural network based at least in part on the one or more gating parameters of the one or more gating units to generate a sparsified neural network; and
deploying the sparsified neural network to perform inference.

### Claim 19 (depends on 18)

The computer-implemented method of claim 18, wherein the one or more gating parameters comprises a gating threshold, wherein sparsifying the neural network comprises dropping one or more network units from the neural network based at least in part on the gating threshold.

### Claim 20 (depends on 18)

The computer-implemented method of claim 18, wherein training the neural network comprises epoch training, wherein a learning rate in subsequent epochs is decremented.

### Claim 21 (depends on 18)

The computer-implemented method claim 18, wherein sparsifying the neural network comprises approximating one or more gating paths based at least in part on the one or more gating parameters.

### Claim 22 (depends on 18)

The computer-implemented method of claim 18, wherein sparsifying the neural network comprises dropping at least one row of a multiplication matrix comprising the neural network.

### Claim 23 (depends on 18)

The computer-implemented method of claim 18, wherein the method further comprises clipping, by the one or more computing devices, one or more outputs of the one or more network units in the sparsified neural network during at least one of a forward propagation step or a backward propagation step after the sparsified neural network is deployed.
