Neural Network Encoder Decoder Optimization via Segmented Cell Types
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods are unable to efficiently optimize the architecture of neural networks with encoders and decoders within a reasonable time using a single GPU, limiting their ability to perform complex tasks like depth determination from stereo images or semantic segmentation.
Innovation Solution
A method that includes providing different cell types connected in a predeterminable order, with reduction and upsampling cells to adjust resolution, and using gradient descent for parameter optimization, allowing for efficient optimization of neural networks with encoders and decoders.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If known optimization methods are used for neural networks with encoders and decoders, then optimization capability is maintained, but computational cost becomes extremely expensive and optimization time becomes unreasonable
Solution Approach 1:
The patent segments the neural network architecture into standardized cell types (encoder cells and decoder cells) with specific functions. Each cell type handles specific operations (downsampling, upsampling, feature extraction), allowing the optimization process to work with modular units rather than the entire complex network at once, thus reducing computational burden while maintaining optimization capability
Solution Approach 2:
The patent changes the parameterization approach by representing cell connections and operations as learnable parameters that can be optimized efficiently. The architecture search space is parameterized such that optimization can proceed with reasonable computational cost, transforming the discrete architecture search into a continuous parameter optimization problem
2Reliability
If complex neural networks with encoders and decoders are optimized, then task performance for demanding tasks improves, but device complexity increases
Solution Approach 1:
The complex neural network is divided into standardized cell types with specific functions (encoder cells for downsampling and feature extraction, decoder cells for upsampling and reconstruction). This segmentation makes the complex architecture manageable and optimizable while maintaining the capability to perform demanding tasks through proper cell composition
Solution Approach 2:
The patent creates universal cell types that can serve multiple functions within the encoder-decoder architecture. These cells are designed to handle various operations (convolution, pooling, upsampling) in a unified manner, reducing the need for specialized complex components while maintaining task performance
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a method (20) for creating a neuronal network (40) which comprises an encoder that is connected to a decoder. The optimization method DARTS is used, wherein a further cell type is added to the known cell-types of DARTS. The invention further relates to a computer program and to a device for carrying out the method (30) as well as to a machine-readable memory element, in which the computer program is stored.