# US 12,217,163 B2 · Methods and systems for budgeted and simplified training of deep neural networks

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

- Number: US 12,217,163 B2
- Document ID: US12217163B2
- Status: Granted
- Kind: B2
- Application number: 18/371,934
- Filed: 2023-09-22
- Published: 2024-03-14
- Granted: 2025-02-04
- Earliest claimed: 2022-01-25
- Assignees: Intel Corporation
- Applicants: Intel Corporation
- Inventors: Yiwen Guo; Yuqing Hou; Anbang Yao; Dongqi Cai; Lin Xu; Ping Hu; Shandong Wang; Wenhua Cheng; Yurong Chen; Libin Wang
- Web page: https://patentgrep.com/patent/US12217163B2

## CPC

- G06N 3/063: using electronic means
- G06F 18/213: Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F 18/2148: {characterised by the process organisation or structure, e.g. boosting cascade}
- G06F 18/217: {Validation; Performance evaluation; Active pattern learning techniques}
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/045: Combinations of networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/08: Learning methods
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 3/09: Supervised learning
- G06N 3/092: Reinforcement learning
- G06N 3/098: Distributed learning, e.g. federated learning
- G06V 10/454: {Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]}
- G06V 10/764: using classification, e.g. of video objects
- G06V 10/82: using neural networks
- G06V 10/94: Hardware or software architectures specially adapted for image or video understanding
- G06V 10/955: {using specific electronic processors}
- G06V 20/00: Scenes; Scene-specific elements (control of digital cameras H04N23/60)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING

## Abstract

Methods and systems for budgeted and simplified training of deep neural networks (DNNs) are disclosed. In one example, a trainer is to train a DNN using a plurality of training sub-images derived from a down-sampled training image. A tester is to test the trained DNN using a plurality of testing sub-images derived from a down-sampled testing image. In another example, in a recurrent deep Q-network (RDQN) having a local attention mechanism located between a convolutional neural network (CNN) and a long-short time memory (LSTM), a plurality of feature maps are generated by the CNN from an input image. Hard-attention is applied by the local attention mechanism to the generated plurality of feature maps by selecting a subset of the generated feature maps. Soft attention is applied by the local attention mechanism to the selected subset of generated feature maps by providing weights to the selected subset of generated feature maps in obtaining weighted feature maps. The weighted feature maps are stored in the LSTM. A Q value is calculated for different actions based on the weighted feature maps stored in the LSTM.

## Claims

### Claim 1

A method for training a deep neural network (DNN), comprising:
receiving one or more input images;
generating a plurality of sub-images from each of the one or more input images;
training the DNN with batches of the plurality of sub-images to obtain a training result by adjusting weights of the DNN based on results from processing the batches of the sub-images.

### Claim 2 (depends on 1)

The method of claim 1, further comprising:
sub-sampling a testing image into a plurality of testing sub-images;
selecting a first testing sub-image of the plurality of testing sub-images; and
testing the trained DNN with the selected first testing sub-image to obtain a first testing result.

### Claim 3 (depends on 2)

The method of claim 2, further comprising:
selecting a second testing sub-image of the plurality of testing sub-images;
testing the trained DNN with the selected second testing sub-image to obtain a second testing result; and
selecting one of the first testing result and second testing result as a testing result.

### Claim 4 (depends on 1)

The method of claim 1, wherein at least one of the plurality of sub-images has a lower resolution than at least one of the one or more input images.

### Claim 5

A system comprising:
a processor to execute a deep neural network (DNN); and
a second processor coupled to the processor, wherein the second processor is to
train the DNN using a plurality of training sub-images derived from a down-sampled training image, and
test the trained DNN using a plurality of testing sub-images derived from a down-sampled testing image.

### Claim 6 (depends on 5)

The system of claim 5, wherein the second processor is to:
sub-sample a testing image into a plurality of testing sub-images,
select a first testing sub-image, and
test the trained DNN with the selected first testing sub-image to obtain a first testing result.

### Claim 7 (depends on 6)

The system of claim 6, wherein the second processor is to:
select a second testing sub-image,
test the trained DNN with the selected second testing sub-image to obtain a second testing result, and
select one of the first testing result and second testing result as a testing result.

### Claim 8 (depends on 5)

The system of claim 5, wherein at least one of the plurality of training sub-images has a lower resolution than the training image.

### Claim 9

A method comprising:
receiving one or more input images;
generating a plurality of sub-images from the one or more input images;
grouping the plurality of sub-images into multiple batches of sub-images; and
training a deep neural network (DNN) with selected batches of sub-images of the multiple batches of sub-images by adjusting weights of the DNN based on results from processing the selected batches of sub-images with the DNN.

### Claim 10 (depends on 9)

The method of claim 9, wherein:
a batch of the multiple batches of sub-images includes sub-images from different input images of the one or more input images.

### Claim 11 (depends on 9)

The method of claim 9, comprising:
providing test sub-images of a test image to the DNN after the training;
processing the test sub-images with the DNN to obtain test results; and
combining the test results to determine a result.

### Claim 12 (depends on 11)

The method of claim 11, wherein:
combining the test results comprises applying a voting method to select the result from the test results.

### Claim 13 (depends on 9)

The method of claim 9, wherein:
generating the plurality of sub-images from the one or more input images includes generating a plurality of lower resolution sub-images.

### Claim 14 (depends on 13)

The method of claim 13, wherein:
generating the plurality of lower resolution sub-images includes down-sampling the one or more input images.