# US 2021/0174200 A1 · Training device and training method for neural network model

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

- Number: US 2021/0174200 A1
- Document ID: US20210174200A1
- Status: Published application
- Kind: A1
- Application number: 16/729,453
- Filed: 2019-12-29
- Published: 2021-06-10
- Earliest claimed: 2019-12-04
- Assignees: Industrial Technology Research Institute
- Applicants: Industrial Technology Research Institute
- Inventors: Mao-Yu Huang; Po-Yen Hsieh; Chih-Neng Liu; Tsann-Tay Tang
- Web page: https://patentgrep.com/patent/US20210174200A1

## CPC

- G06N 3/08: Learning methods
- G06T 7/0002: {Inspection of images, e.g. flaw detection}
- G01N 21/956: Inspecting patterns on the surface of objects {(contactless testing of electronic circuits G01R31/308; testing currency G07D; manufacturing processes per se of semiconductor devices implementing a measuring step H10P74/20)}
- G06N 3/045: Combinations of networks
- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/0499: Feedforward networks
- G06N 3/09: Supervised learning
- G06T 7/001: {using an image reference approach}
- G01N 2021/8883: {involving the calculation of gauges, generating models}
- G01N 2021/95646: {Soldering}
- G01N 2201/1296: using neural networks
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G06T 2207/20081: Training; Learning
- G06T 2207/20084: Artificial neural networks [ANN]
- G06T 2207/30141: Printed circuit board [PCB]
- G06T 2207/30148: Semiconductor; IC; Wafer
- G06T 2207/30152: Solder
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06T: IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G01N: INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES (measuring or testing apparatus or processes other than immunoassay, involving enzymes or microorganisms C12M, C12Q)

## Abstract

A training device and a training method for a neural network model are provided. The training method includes: obtaining a data set; completing, according to the data set, a plurality of artificial intelligence (AI) model trainings to generate a plurality of models corresponding to the plurality of AI model trainings respectively; selecting, according to a first constraint, a first model set from the plurality of models; and selecting, according to a second constraint, the neural network model from the first model set.

## Claims

### Claim 1

A training device for a neural network model, comprising:
a transceiver;
a storage medium, storing a plurality of modules; and
a processor, coupled to the storage medium and the transceiver, and accessing and executing the plurality of modules, wherein the plurality of modules comprise:
a data collection module, obtaining a data set through the transceiver; and
a training module, completing a plurality of artificial intelligence (AI) model trainings according to the data set to generate a plurality of models corresponding to the plurality of AI model trainings respectively, selecting a first model set from the plurality of models according to a first constraint, and selecting the neural network model from the first model set according to a second constraint.

### Claim 2 (depends on 1)

The training device as claimed in claim 1, wherein the training module selects a second model set from the first model set according to a third constraint and selects the neural network model from the second model set according to the second constraint.

### Claim 3 (depends on 1)

The training device as claimed in claim 1, wherein the training module selects the neural network model corresponding to a target condition from the first model set according to the second constraint.

### Claim 4 (depends on 1)

The training device as claimed in claim 1, wherein the training module selects the neural network model corresponding to the last iteration among the iterations of AI model training complied with the second constraint from the first model set according to the second constraint.

### Claim 5 (depends on 1)

The training device as claimed in claim 1, wherein the first constraint and the second constraint are associated with at least one of: a true negative rate (TNR), a true positive rate (TPR), a false negative rate (FNR), a false positive rate (FPR), a minimum error rate, an average error rate, a recall rate, or accuracy.

### Claim 6 (depends on 1)

The training device as claimed in claim 1, wherein the neural network model comprises an autoencoder neural network, a deep learning neural network, a deep residual learning neural network, and a restricted Boltzmann machine neural network, a recursive neural network, or a multilayer perception neural network.

### Claim 7 (depends on 1)

The training device as claimed in claim 1, wherein the data collection module receives defect samples and normal samples through the transceiver and oversamples the defect samples to generate the data set according to the oversampled defect samples and the normal samples.

### Claim 8 (depends on 7)

The training device as claimed in claim 7, wherein the data set is associated with an appearance of a printed circuit board, and the defect samples are associated with at least one of: inclusion of solder mask (S/M), scratch of S/M, pollution of S/M, copper exposure of S/M, uneven printing of S/M, skip printing of S/M, poor developing, board damage of S/M, wetting of S/M, inclusion of S/M organic solderability preservative (OSP), oxidation of S/M OSP, pollution of PAD gold plating, oxidation of PAD gold plating, copper exposure of PAD gold plating, blur text, wetting of PAD or normal.

### Claim 9 (depends on 7)

The training device as claimed in claim 7, wherein the data set is associated with an appearance of a semiconductor manufacturing wafer, and the defect samples are associated with at least one of: scratch, arrow, particle, discolor, or normal.

### Claim 10 (depends on 7)

The training device as claimed in claim 7, wherein the data collection module oversamples the defect samples in response to a ratio of the defect samples to the normal samples being less than a threshold.

### Claim 11 (depends on 1)

The training device as claimed in claim 1, further comprising:
a graphical user interface, coupled to the processor, wherein the processor obtains the first constraint and the second constraint through the graphical user interface.

### Claim 12

A training method for a neural network model, comprising:
obtaining a data set;
completing a plurality of artificial intelligence (AI) model trainings according to the data set to generate a plurality of models corresponding to the plurality of AI model trainings respectively;
selecting a first model set from the plurality of models according to a first constraint; and
selecting the neural network model from the first model set according to a second constraint.

### Claim 13 (depends on 12)

The training method as claimed in claim 12, wherein the step of selecting the neural network model from the first model set according to the second constraint comprises:
selecting a second model set from the first model set according to a third constraint and selecting the neural network model from the second model set according to the second constraint.

### Claim 14 (depends on 12)

The training method as claimed in claim 12, wherein the step of selecting the neural network model from the first model set according to the second constraint comprises:
selecting the neural network model corresponding to a target condition from the first model set according to the second constraint.

### Claim 15 (depends on 12)

The training method as claimed in claim 12, wherein the step of selecting the neural network model from the first model set according to the second constraint comprises:
selecting the neural network model corresponding to the last iteration among the iterations of AI model training complied with the second constraint from the first model set according to the second constraint.

### Claim 16 (depends on 12)

The training method as claimed in claim 12, wherein the first constraint and the second constraint are associated with at least one of: a true negative rate (TNR), a true positive rate (TPR), a false negative rate (FNR), a false positive rate (FPR), a minimum error rate, an average error rate, a recall rate, or accuracy.

### Claim 17 (depends on 12)

The training method as claimed in claim 12, wherein the neural network model comprises an autoencoder neural network, a deep learning neural network, a deep residual learning neural network, and a restricted Boltzmann machine neural network, a recursive neural network, or a multilayer perception neural network.

### Claim 18 (depends on 12)

The training method as claimed in claim 12, wherein the step of obtaining the data set comprises:
receiving defect samples and normal samples and oversampling the defect samples to generate the data set according to the defect samples and the normal samples.

### Claim 19 (depends on 18)

The training method as claimed in claim 18, wherein the data set is associated with an appearance of a printed circuit board, and the defect samples are associated with at least one of: inclusion of solder mask (S/M), scratch of S/M, pollution of S/M, copper exposure of S/M, uneven printing of S/M, skip printing of S/M, poor developing, board damage of S/M, wetting of S/M, inclusion of S/M organic solderability preservative (OSP), oxidation of S/M OSP, pollution of PAD gold plating, oxidation of PAD gold plating, copper exposure of PAD gold plating, blur text, wetting of PAD, or normal.

### Claim 20 (depends on 18)

The training method as claimed in claim 18, wherein the data set is associated with an appearance of a semiconductor manufacturing wafer, and the defect samples are associated with at least one of: scratch, arrow, particle, discolor, or normal.

### Claim 21 (depends on 18)

The training method as claimed in claim 18, wherein the data collection module oversamples the defect samples in response to a ratio of the defect samples to the normal samples being less than a threshold.

### Claim 22 (depends on 12)

The training method as claimed in claim 12, further comprising:
obtaining the first constraint and the second constraint through a graphical user interface.
