# US 2024/0430136 A1 · Forward-forward learning based wireless communications systems

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

- Number: US 2024/0430136 A1
- Document ID: US20240430136A1
- Status: Published application
- Kind: A1
- Application number: 18/734,087
- Filed: 2024-06-05
- Published: 2024-12-26
- Earliest claimed: 2023-06-26
- Assignees: Apple Inc.
- Applicants: Apple Inc.
- Inventors: Zhou Zhou; Sharad Sambhwani
- Web page: https://patentgrep.com/patent/US20240430136A1

## CPC

- H04L 25/0254: {using neural network algorithms}
- H04L 25/0252: {using third or higher order statistics}
- H04L: TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION (arrangements common to telegraphic and telephonic communication H04M)

## Abstract

A wireless communication system may use forward-forward learning to for end-to-end learning. A transmitter may pass a positive dataset and a negative dataset through each of its layers for model training. Each layer may correspond to a goodness function. The transmitter may send the positive dataset to a receiver. The receiver may generate a second positive dataset and a second negative dataset based on the positive dataset sent from the receiver. The receiver may train each of its layers using the second positive dataset and the second negative dataset.

## Claims

### Claim 1

A method for training an end-to-end learning based wireless communication system, the method comprising:
generating, at a transmitter, a first positive dataset and a first negative dataset;
training one or more layers of the transmitter using the first positive dataset and the first negative dataset;
sending, via the transmitter, the first positive dataset as signals over a channel to a receiver;
generating, at the receiver, a second positive dataset and a second negative dataset from the first positive dataset; and
training one or more layers of the receiver using the second positive dataset and the second negative dataset.

### Claim 2 (depends on 1)

The method of claim 1, wherein the second positive dataset and the second negative dataset is created by combining the received signals from the channel with an auxiliary input.

### Claim 3 (depends on 1)

The method of claim 1, wherein training the one or more layers of the transmitter uses forward-forward learning by passing the first positive dataset and the first negative dataset through each layer of the transmitter,
wherein the one or more layers of the receiver use forward-forward learning by passing the second positive dataset and the second negative dataset through each layer of the receiver,
wherein each layer of the receiver and the transmitter corresponds to its own individual goodness function, and
wherein the transmitter and the receiver maximize the individual goodness functions for each layer when the first positive dataset and the second positive dataset are fed through the one or more layers of the transmitter and the one or more layers of the receiver respectively, and
wherein the transmitter and the receiver minimize the individual goodness functions for each layer when the first negative dataset and the second negative dataset are fed through the one or more layers of the transmitter and the one or more layers of the receiver respectively.

### Claim 4 (depends on 1)

The method of claim 1, wherein each of the one or more layers of the transmitter and each of the one or more layers of the receiver are trained without storing gradient information of any other layers.

### Claim 5 (depends on 1)

The method of claim 1, wherein each of the one or more layers of the transmitter and each of the one or more layers of the receiver are trained following a same order of a communication process.

### Claim 6 (depends on 1)

The method of claim 1, wherein each of the one or more layers of the transmitter and each of the one or more layers of the receiver are trained without obtaining a model of the channel.

### Claim 7 (depends on 1)

The method of claim 1, wherein training of the one or more layers of the receiver is based on measuring activations of the one or more layers of the receiver.

### Claim 8 (depends on 1)

The method of claim 1, further comprising altering a number of layers to be trained based on a wireless communication link metric.

### Claim 9 (depends on 1)

The method of claim 1, further comprising adding, removing, freezing, or activating one or more of the one or more layers of the transmitter or the one or more layers of the receiver.

### Claim 10 (depends on 1)

The method of claim 1, further comprising:
sending, via the receiver, a feedback signal to the transmitter; and
combining the feedback signal with a targeting layer at the transmitter, wherein the feedback signal can be combined with any layer of the transmitter, wherein the feedback signal is constructed by applying a fixed linear mapping to activations of the receiver.

### Claim 11 (depends on 1)

The method of claim 1, wherein the transmitter can send the first positive dataset based on a lookup table that maps a group-of-bits to a label.

### Claim 12 (depends on 1)

The method of claim 1, further comprising using a tree-structure dividing process to separate the second positive dataset and the second negative dataset into a plurality of sub-blocks, wherein the training is done using the sub-blocks.

### Claim 13 (depends on 1)

The method of claim 1, further comprising inferring, at the receiver, a value sent by the transmitter based on goodness functions associated with each of the one or more layers of the receiver.

### Claim 14

A transmitter apparatus comprising:
a processor; and
a memory storing instructions that, when executed by the processor, configure the apparatus to:
generate a first positive dataset and a first negative dataset;
train one or more layers of the transmitter using the first positive dataset and the first negative dataset; and
send the first positive dataset as signals over a channel to a receiver.

### Claim 15 (depends on 14)

The transmitter apparatus of claim 14, wherein training the one or more layers of the transmitter apparatus uses forward-forward learning by passing the first positive dataset and the first negative dataset through each layer of the transmitter,
wherein each layer of the transmitter apparatus corresponds to its own individual goodness function, and
wherein the transmitter apparatus maximizes the individual goodness functions for each layer when the first positive dataset is fed through the one or more layers of the transmitter apparatus, and
wherein the transmitter apparatus minimizes the individual goodness functions for each layer when the first negative dataset is fed through the one or more layers of the transmitter apparatus.

### Claim 16 (depends on 14)

The transmitter apparatus of claim 14, wherein each of the one or more layers of the transmitter apparatus are trained without storing gradient information of any other layers.

### Claim 17 (depends on 14)

The transmitter apparatus of claim 14, wherein each of the one or more layers of the transmitter apparatus are trained follow a same order of a communication process.

### Claim 18 (depends on 14)

The transmitter apparatus of claim 14, wherein each of the one or more layers of the transmitter apparatus are trained without obtaining a model of the channel.

### Claim 19 (depends on 14)

The transmitter apparatus of claim 14, wherein training of the one or more layers of the transmitter is based on measuring activations of the layers.

### Claim 20 (depends on 14)

The transmitter apparatus of claim 14, wherein the instructions further configure the apparatus to alter a number of layers to be trained based on a wireless communication link metric.
