US-12625923-B2 - System and method for adjusting input data of neural network

US12625923B2US 12625923 B2US12625923 B2US 12625923B2US-12625923-B2

Abstract

A system and method for adjusting input data of a decision-making neural network is provided, wherein the system includes a data-dividing neural network apparatus and a data processing apparatus. The data-dividing neural network apparatus receives an input data and divides the input data into a plurality of sub data including a first sub data and a second sub data. The data processing apparatus is coupled to the data-dividing neural network apparatus to receive the sub data, and process the first sub data and the second sub data by different ways when the sub data is processed, so that the first sub data and the second sub data are differently adjusted. The decision-making neural network is electrically coupled to the data processing apparatus to take the processed sub data as input data. As a result, the neural network can change the final output results.

Inventors

Jia-yo HSU
I-CHIH CHEN

Assignees

VIA TECHNOLOGIES, INC.

Dates

Publication Date: 20260512
Application Date: 20210326
Priority Date: 20201123

Claims (18)

1 . A system for adjusting input data of a neural network, adapted to adjust input data prior to inference by a decision-making neural network in order to improve inference stability and robustness, the system comprising: a feature-partitioning neural network circuit receiving multi-dimensional input data representing a plurality of features and configured to generate a plurality of sub data by partitioning the input data into feature groups based on learned feature correlation values, the plurality of sub data including at least a first sub data representing a first feature group and a second sub data representing a second feature group; a data processing circuit coupled to the feature-partitioning neural network circuit and configured to: apply a first transformation function to the first sub data, apply a second, different transformation function to the second sub data, wherein the first and second transformation functions apply different normalization, weighting, or scaling operations selected to modify a correlation magnitude between the first sub data and the second sub data, and combine the transformed first sub data and transformed second sub data into modified input data having an adjusted inter-feature relationship; wherein the decision-making neural network is coupled to receive the modified input data and is configured to generate a decision output based on the adjusted inter-feature relationship, and wherein the data processing circuit modifies the inter-feature relationship prior to inference to reduce sensitivity of the decision-making neural network to noise or bias present in the original input data.
2 . The system of claim 1 , wherein the input data specifies image data represented as a plurality of image features.
3 . The system of claim 2 , wherein each of the plurality of sub data specifies a feature group corresponding to a partial image region of the image; and after the first sub data and the second sub data are processed by the data processing apparatus, a computed distance metric between a first partial image region represented by the first sub data and a second partial image region represented by the second sub data is changed.
4 . The system of claim 3 , wherein the decision-making neural network is configured to determine the result according to the computed distance metric between the first partial image region and the second partial image region.
5 . The system of claim 1 , wherein the data processing circuit is configured to modify the inter-feature relationship between the first sub data and the second sub data by different transformation magnitudes to generate a plurality of different modified input data; and wherein the plurality of different modified input data is respectively input to a plurality of decision-making neural networks to generate a plurality of different results, each decision-making neural being configured to determine a result according to the inter-feature relationship corresponding to the modified input data received by that decision-making neural network.
6 . The system of claim 3 , wherein a first partial image region represented by the first sub data and a second partial image region represented by the second sub data correspond to non-overlapping feature groups.
7 . The system of claim 1 , wherein the data-dividing neural network circuit comprises a semantic analysis neural network configured to perform feature partitioning of the input data.
8 . The system of claim 1 , wherein the decision-making neural network is configured to initially receive the input data to generate a first result; and wherein, in response to the first result failing to satisfy a predefined machine-evaluated standard, the data processing circuit is configured to generate the modified input data, and the decision-making neural network is configured to receive the modified input data to perform a subsequent inference operation to generate an updated result.
9 . The system of claim 8 , wherein the data processing circuit is configured to determine how the inter-feature relationship between the first sub data and the second sub data is to be modified based on the first result generated by the decision-making neural network.
10 . A method carried out by a processor executing program code for adjusting input data of a neural network prior to inference by a decision-making neural network, the method comprising: receiving, by a processor, multi-dimensional input data representing a plurality of features; partitioning, using a feature-partitioning neural network, the input data into a plurality of sub data including at least a first sub data representing a first feature group and a second sub data representing a second feature group, the partitioning based on learned feature correlation values; processing, by the processor, the first sub data using a first transformation function and the second sub data using a second transformation function different from the first transformation function, such that an inter-feature correlation between the first sub data and the second sub data is modified; combining the processed first sub data and processed second sub data into modified input data; and inputting the modified input data to the decision-making neural network to perform an inference operation, wherein the decision-making neural network determines a result based on the modified inter-feature correlation.
11 . The method of claim 10 , wherein the input data specifies image data represented as a plurality of image features.
12 . The method of claim 11 , wherein each of the plurality of sub data comprises a feature group corresponding to a partial image region of the image; and after the first sub data and the second sub data are processed using different transformation functions, a computed distance metric between a first partial image region represented by the first sub data and a second partial image region represented by the second sub data is changed.
13 . The method of claim 12 , wherein the decision-making neural network determines the result based on the computed distance metric between the first partial image region and the second partial image region.
14 . The method of claim 10 , wherein processing the first sub data and the second sub data using different transformation functions comprises: modifying an inter-feature relationship between the first sub data and the second sub data by different transformation magnitudes to generate a plurality of different modified input data, respectively; inputting the plurality of different modified input data respectively to a plurality of decision-making neural networks to generate a plurality of different results, each decision-making neural network determining a result according to the inter-feature relationship corresponding to the modified input data received by that decision-making neural network.
15 . The method of claim 12 , wherein a first partial image region represented by the first sub data and a second partial image region represented by the second sub data correspond to non-overlapping feature groups.
16 . The method of claim 10 , wherein dividing the input data into a plurality of sub data comprises using a semantic analysis neural network configured to perform feature partitioning to generate the plurality of sub data.
17 . The method of claim 10 , wherein before processing the first sub data and the second sub data using different transformation functions, the method further comprises: inputting the input data to the decision-making neural network to perform an initial inference operation and generate a first result; and in response to the first result failing to satisfy a predefined machine-evaluated standard, processing the first sub data and the second sub data using the different transformation functions to generate the modified input data and performing a subsequent inference operation using the modified input data.
18 . The method of claim 17 , further comprising: determining how an inter-feature relationship between the first sub data and the second sub data is to be modified based on the first result generated by the decision-making neural network.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS This application claims the priority benefit of Taiwan application serial no. 109140945, filed on Nov. 23, 2020. The entirety of the above-mentioned patent application is hereby incorporated by reference herein and made a part of this specification. TECHNICAL FIELD The present disclosure relates to a system for assisting a neural network, in particular to a system and method for adjusting input data of a neural network. BACKGROUND Neural network is a computer application that performs calculations by connections of a large number of artificial neurons. In most cases, the neural network can gradually change the internal structure based on external information, therefore to some extent having a learning function. After the neural network is built, trainings thereof can be performed by accumulating external information, and the structure of itself can be improved gradually; after the training is completed, determined results of substantial accuracy can be provided. When a neural network can provide determined results of substantial accuracy, it means that for same external input condition, only same results are provided by the same neural network. From another point of view, if the neural network is required to generate other results for the same external input condition, then the neural network must be trained for this goal. However, each time of the process of self-training by accumulating external information costs a lot of time. Therefore, when facing an operating procedure where the determining conditions may change frequently, it may not be worthy to introduce a neural network. For example, assuming that for packaging seals of a product A, to meet a requirement, only a gap smaller than 1 mm between adjacent seals is required, while for packaging seals of a product B, to meet a requirement, a gap smaller than 0.5 mm between adjacent seals is required, then a computer vision neural network for checking the packaging seals of product A cannot be shared as a computer vision neural network for checking the packaging seals of product B. In this situation, by technologies in prior art, the problem can only be solved by establishing two sets of computer vision neural networks corresponding to the seal checks of the two different products respectively. Once other products with different requirements on the packaging seals appear, the problem can only be solved by increasing the number of computer vision neural networks or re-training the computer vision neural networks that are no longer in use. Therefore, although a neural network can provide accurate determined results, since only accurate determined results can be provided, a neural network may not be suitable for solving problems in some special environments. SUMMARY In view of this, the description of the present disclosure provides a system and a method for adjusting the input data of a neural network, wherein the neural network is caused to change the final output results by changing the contents of data that was to be input to the neural network. In one aspect, the present disclosure provides a system for adjusting input data of a neural network; the system is adapted to adjust input data which is to be input to a decision-making neural network and includes a data-dividing neural network apparatus and a data processing apparatus. The data-dividing neural network apparatus is configured to receive the input data and divide the input data into a plurality of sub data including a first sub data and a second sub data; the data processing apparatus is coupled to the data-dividing neural network apparatus to receive each sub data, and configured to process the first sub data and the second sub data by different ways, so that the first sub data and the second sub data are differently adjusted, and combine the first adjusted sub data and the second adjusted sub data as modified input data; wherein the decision-making neural network is coupled to the data processing apparatus to take the modified input data as the input data. The decision-making neural network is configured to determine a result based on a relationship between the first sub data and the second sub data which may have been changed by the data processing apparatus. In one embodiment, the input data is an image. In one embodiment, each sub data is a partial image of the image; after the first sub data and the second sub data are processed by the data processing apparatus, a distance between a first partial image represented by the first sub data and a second partial image represented by the second sub data is changed. Further, the decision-making neural network is configured to determine the result according to the distance between the first partial image and the second partial image. In another embodiment, the data processing apparatus is configured to change the relationship between the first sub data and the second sub data by different degrees to combine th