# US 12,182,707 B2 · System and method for the discovering efficient random neural networks

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

- Number: US 12,182,707 B2
- Document ID: US12182707B2
- Status: Granted
- Kind: B2
- Application number: 18/272,747
- Filed: 2022-02-18
- Published: 2024-03-21
- Granted: 2024-12-31
- Earliest claimed: 2021-02-18
- Assignees: Carnegie Mellon University
- Applicants: Carnegie Mellon University
- Inventors: Marios Savvides; Thanh Hai Phan; Zhiqiang Shen; Akshay Chawla
- Web page: https://patentgrep.com/patent/US12182707B2

## CPC

- G06N 3/04: Architecture, e.g. interconnection topology
- G06N 3/08: Learning methods
- G06N 3/006: based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/0495: Quantised networks; Sparse networks; Compressed networks
- G06N 3/082: modifying the architecture, e.g. adding, deleting or silencing nodes or connections
- G06N 3/09: Supervised learning
- G06N 3/092: Reinforcement learning
- G06N 3/0985: Hyperparameter optimisation; Meta-learning; Learning-to-learn
- G06N 7/01: Probabilistic graphical models, e.g. probabilistic networks
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

Disclosed herein is a system and method for novel neural architecture search using a random graph network backbone to facilitate the creation of an efficient network structure. The method utilizes reinforcement learning algorithms to build a complex relationship between intra-connections (i.e., links between blocks in a random graph network) and extra-connections (i.e., links among blocks across the random graphs network) for discovering an efficient random neural architecture.

## Claims

### Claim 1

A method for optimizing a graph network representing a neural network comprising:
training the graph network using a deep deterministic policy gradient, resulting in a chain of random graph blocks, each random graph block defining a state of the neural network;
wherein the deep deterministic policy gradient defines an actor network to propose an action based on the state, the action comprising parameters defining a number of nodes in each random graph block and both intra-block connections and extra-block connections; and
wherein the deep deterministic policy gradient defines a critic network that evaluates the proposed actions of the actor function.

### Claim 2 (depends on 1)

The method of claim 1 wherein:
each random graph block comprises a plurality of nodes connected by a plurality of edges;
intra-block connections define connections between nodes in a single block; and
extra-block connections define connections between nodes in different blocks.

### Claim 3 (depends on 2)

The method of claim 2 wherein the parameters defining the intra-block connections comprise:
k, an average degree over the block; and
p, a probability that two arbitrary nodes in the block are connected.

### Claim 4 (depends on 3)

The method of claim 3 wherein the parameters defining extra-block connections define extra-block connections for the current block and each subsequent block in the graph network.

### Claim 5 (depends on 2)

The method of claim 2 wherein the parameters defining the intra-block and extra-block connections comprise:
k′, an average degree over the block and each subsequent block in the graph; and
p′, a probability that an arbitrary node in one block is connected to an arbitrary node in a subsequent block.

### Claim 6 (depends on 5)

The method of claim 5 wherein each node in a block is connected to k/2 other nodes.

### Claim 7 (depends on 6)

The method of claim 6 wherein an action space of the actor function is defined as: a=(N, k, p, k′, p′);
wherein:
N is the number of nodes in the block.

### Claim 8 (depends on 7)

The method of claim 7 wherein a state space of the graph network is defined as: s=(, N, C, k, p, k′, p′);
wherein:
denotes a kernel size of a depth-wise convolution; and
C is a number of output channels.

### Claim 9 (depends on 1)

The method of claim 1 wherein the final state of the graph network is achieved when a loss of the critic network is minimized.

### Claim 10 (depends on 9)

The method of claim 9 wherein, when the final status graph network is been achieved, implementing the neural network based on the final state of the graph network.

### Claim 11 (depends on 1)

A system comprising:
a processor; and
memory, storing software that, when executed by a processor, implements the method of claim 1.

### Claim 12 (depends on 10)

A system comprising:
a processor; and
memory, storing software that, when executed by a processor, implements the method of claim 10.