IP Library › Granted Patent US 12,639,555
Granted Patent B2
US 12,639,555 · App. 19/204,647 · Granted May 26, 2026

Method and device for optimizing neural network model

Inventors: Jungi Lee (Seoul, KR); Ji Seong Yoon (Seoul, KR); Myoung Hwan Kim (Seoul, KR); Jeong Hyeon Park (Seoul, KR); Kwangsun Yoo (Incheon, KR); Seok Joo Byun (Gyeonggi-do, KR)
Assignee: EL ROI LAB INC.
G06N3/048G06N3/0495
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,639,555
App. No.
19/204,647
Granted
May 26, 2026
Kind
B2
Abstract

A method of optimizing a neural network model, performed by a neural network model optimization device, may include removing an activation function included in one of a consecutive first block and second block, fusing batch normalization function and a fully connected layer included in a block included in the neural network model, and fusing the first block and the second block.

Claims (26)

1 . A method of optimizing a neural network model, performed by a neural network model optimization device, the method comprising:

removing an activation function included in one of a consecutive first block and second block;

fusing batch normalization function and a fully connected layer included in a block included in the neural network model; and

fusing the first block and the second block,

wherein the fusing the first block and the second block comprises fusing the first block and the second block if an activation function is absent in the first block and an activation function is present in the second block,

wherein the fusing of the first block and the second block comprises fusing a first layer included in the first block and a second layer included in the second block, and

a third block generated by fusing the first block and the second block includes a third layer in which the first layer and the second layer are fused and an activation function included in the second block,

wherein the method further comprising:

reducing the number of hidden layers included in the neural network model based on a hidden layer order,

wherein the hidden layer order is numerical value assigned to each hidden layer with higher value assigned as it gets closer to input of the neural network model and lower value assigned as it gets closer to output of the neural network model,

wherein the number of hidden layers is adjusted such that the number of hidden layers corresponds to square of the hidden layer order.

2 . The method of claim 1 , wherein the fusing of the fully connected layer and the batch normalization comprises generating a fusion layer by fusing the fully connected layer and the batch normalization function.

3 . A neural network model optimization device comprising:

a memory configured to store a neural network model optimization program for optimizing a neural network model; and

a processor configured to control the memory,

wherein the processor is configured to,

remove an activation function included in one of a consecutive first block and second block,

fuse batch normalization function and a fully connected layer included in a block included in the neural network model, and

fuse the first block and the second block if an activation function is absent in the first block and an activation function is present in the second block,

wherein the processor is further configured to,

fuse a first layer included in the first block and a second layer included in the second block, and

a third block generated by fusing the first block and the second block includes a third layer in which the first layer and the second layer are fused and an activation function included in the second block,

wherein the processor is further configured to,

reduce the number of hidden layers included in the neural network model based on a hidden layer order,

wherein the hidden layer order is numerical value assigned to each hidden layer with higher value assigned as it gets closer to input of the neural network model and lower value assigned as it gets closer to output of the neural network model,

wherein the number of hidden layers is adjusted such that the number of hidden layers corresponds to square of the hidden layer order.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Aug 11, 2025
From: LEE, JUNGI; YOON, JI SEONG; KIM, MYOUNG HWAN; PARK, JEONG HYEON; YOO, KWANGSUN; BYUN, SEOK JOO
To: EL ROI LAB INC.
Reel/Frame 071978/0646 →
Priority Claims (1)
KR 10-2024-0061725 · May 10, 2024 · national
Continuity (1)
Related Publication 20250348714A1 · Nov 13, 2025
References Cited (14)
US 20210150347A1 · Borse · 2021 [cited by examiner]
US 20210182685A1 · Xiao · 2021 [cited by examiner]
US 20220327386A1 · Ben-dror · 2022 [cited by examiner]
KR 1020220139248A · 2022 [cited by applicant]
Montalbo, Francis Jesmar P. “Fusing compressed deep ConvNets with a self-normalizing residual block and alpha dropout for a cost-efficient classification and diagnosis of gastrointestinal tract diseases.” MethodsX 9 (20… [cited by examiner]
Yu, Xiaohong, et al. “Improving the performance of convolutional neural networks by fusing low-level features with different scales in the preceding stage.” IEEE Access 9 (2021): 70273-70285. (Year: 2021). [cited by examiner]
Nguyen, Dang, et al. “On cross-layer alignment for model fusion of heterogeneous neural networks.” ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023. (Year: 2… [cited by examiner]
O'Neill, James, Greg Ver Steeg, and Aram Galstyan. “Compressing Deep Neural Networks via Layer Fusion.” (2020). (Year: 2020). [cited by examiner]
Singh, Sidak Pal, and Martin Jaggi. “Model fusion via optimal transport.” Advances in Neural Information Processing Systems 33 (2020): 22045-22055. (Year: 2020). [cited by examiner]
Chitty-Venkata et al. “Calibration data-based CNN filter pruning for efficient layer fusion.” 2020 IEEE 22nd International Conference on High Performance Computing and Communications; IEEE 18th International Conf on Sma… [cited by examiner]
Amir Ben Dror, “Layer Folding: Neural Network Depth Reduction using Activation Linearization”, arXiv preprint arXiv:2106.09309, Sep. 3, 2021. [cited by applicant]
Ji Liu, “UPDP: A Unified Progressive Depth Pruner for CNN and VisionTransformer”, arXiv:2401.06426, Jan. 12, 2024. [cited by applicant]
KIPO, Decision to Grant a Patent for Korean Patent Application No. 10-2024-0061725, Jan. 22, 2025. [cited by applicant]
KIPO, Notice of Refusal to Grant a Patent for Korean Patent Application No. 10-2024-0061725, Jul. 10, 2024. [cited by applicant]