Object recognition with reduced neural network weight precision

• US 10,417,525 B2
• Filed: 03/19/2015
• Issued: 09/17/2019
• Est. Priority Date: 09/22/2014
• Status: Active Grant

• Current Assignee: Samsung Electronics Co. Ltd.
• Original Assignee: Samsung Electronics Co. Ltd.
• Inventors: Ji, Zhengping, Ovsiannikov, Ilia, Wang, Yibing Michelle, Shi, Lilong
• Primary Examiner(s): Beg, Samah A
• Application Number:
  US14/663,233
• Publication Number:
  US 20160086078A1
• Time in Patent Office: 1,643 Days
• Field of Search: None
• CPC Class Codes: G06F 18/213 Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
• CPC Class Codes: G06F 18/24 Classification techniques
• CPC Class Codes: G06F 18/24137 Distances to cluster centroïds
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent
• CPC Class Codes: G06V 10/454 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
• CPC Class Codes: G06V 10/7715 Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods
• CPC Class Codes: G06V 10/82 using neural networks
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/18057 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
• CPC Class Codes: G06V 30/19127 Extracting features by transforming the feature space, e.g. multidimensional scaling; Mappings, e.g. subspace methods
• CPC Class Codes: G06V 30/19173 Classification techniques

Abstract

A client device configured with a neural network includes a processor, a memory, a user interface, a communications interface, a power supply and an input device, wherein the memory includes a trained neural network received from a server system that has trained and configured the neural network for the client device. A server system and a method of training a neural network are disclosed.

Claims

1. A client device configured with a trained neural network, the client device comprising:

a processor, a memory, a user interface, a communications interface, a power supply and an input device;

the memory comprising the trained neural network received from a server system, wherein the server system has trained and configured a server-based neural network to be used as the trained neural network for the client device;

wherein;

the trained neural network is configured to generate a feature map, the feature map comprising a plurality of weight values derived from an input image; and

the trained neural network is configured to perform a unitary quantizing operation or a supervised iterative quantization operation on the feature map to reduce a number of bits of each weight of the plurality of weight values from a first predetermined number to a second predetermined number that is less than the first predetermined number without changing a dimension of the feature map.

2. The client device as in claim 1, wherein the input device is configured to capture an image and to store image input data in the memory.

3. The client device as in claim 1, further comprising a multilayer perceptron (MLP) classifier configured to map image input data.

4. The client device as in claim 1, wherein the trained neural network comprises a convolutional neural network.

5. The client device as in claim 1, wherein the quantization operation performs back-propagation (BP) of image input data.

6. The client device as in claim 1, wherein the trained neural network is configured to perform object recognition.

7. The client device as in claim 1, comprising one of a smartphone, a tablet computer and a portable electronic device.

8. The client device as in claim 1, wherein the trained neural network is a quantized low-bit version of the server-based neural network.

9. The client device as in claim 1, wherein network weights for the trained neural network are quantized for lower bit resolution by the server system.

10. A method that comprises performing the following using a client device:

receiving a trained neural network from a server system, wherein the server system has trained and configured a server-based neural network to be used as the trained neural network for the client device;

capturing an input image;

processing the input image using the trained neural network;

generating a feature map using the trained neural network, the feature map comprising a plurality of weight values derived from an input image;

performing a unitary quantizing operation or a supervised iterative quantization operation on the feature map using the trained neural network to reduce a number of bits of each weight of the plurality of weight values from a first predetermined number to a second predetermined number that is less than the first predetermined number without changing a dimension of the feature map; and

recognizing an object in the input image based on a result of the processing.

11. The method of claim 10, wherein receiving the trained neural network includes:

receiving a quantized low-bit version of the server-based neural network as the trained neural network.

12. The method of claim 11, wherein receiving the quantized low-bit version includes:

receiving network weights associated with the trained neural network, wherein the network weights are quantized for lower bit resolution by the server system.

13. The method of claim 10, wherein recognizing the object includes:

analyzing the result of the processing using a multilayer perceptron (MLP) classifier to recognize the object in the input image.

14. A non-transitory computer-readable medium storing program code, which, when executed by a processor, cause the processor to perform the following:

receive a trained neural network from a server system, wherein the server system has trained and configured a server-based neural network to be used as the trained neural network;

capture an input image;

generating a feature map using the trained neural network, the feature map comprising a plurality of weight values derived from an input image;

performing a unitary quantizing operation or supervised iterative quantization operation on the feature map using the trained neural network to reduce a number of bits of each weight of the plurality of weight values from a first predetermined number to a second predetermined number that is less than the first predetermined number without changing a dimension of the feature map; and

process the input image using the trained neural network to recognize an object in the input image.

15. The non-transitory computer-readable medium of claim 14, wherein the program code, when executed by the processor, cause the processor to further perform the following:

receive network weights associated with the trained neural network, wherein the network weights are quantized for lower bit resolution by the server system.

16. The non-transitory computer-readable medium of claim 14, wherein the program code, when executed by the processor, cause the processor to further perform the following:

importing a low-resolution configuration of the server-based neural network; and

storing the imported configuration as the trained neural network.

17. A client device configured with a trained neural network, the client device comprising:

a processor, a memory, a user interface, a communications interface, a power supply and an input device;

the memory comprising the trained neural network received from a server system, wherein the server system has trained and configured a server-based neural network to be used as the trained neural network for the client device;

wherein;

the trained neural network is configured to generate a feature map, the feature map comprising a plurality of first weight values derived from an input image;

the trained neural network is configured to convert the first weights of the feature map into second weights by a unitary or a supervised iterative quantizing operation; and

the second weights are encoded using a number of bits lower than that used to encode the first weights without changing a dimension of the feature map.