Dual Encoder-Decoder Neural Network for Unsupervised Pose Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Encoder-decoder neural networks trained without supervision often lack sufficient accuracy in encoding and decoding data, leading to poor performance in tasks such as human pose estimation and annotation.
Innovation Solution
A dual network system is employed, where one encoder-decoder neural network learns to encode actual attribute values and another learns to calculate differences from attribute averages, with a predefined data structure for the latent space, allowing for unsupervised training and improved accuracy by comparing outputs and adjusting parameters through backpropagation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If encoder-decoder neural networks are trained without supervision, then training cost and time are reduced, but accuracy and performance are insufficient
Solution Approach 1:
The patent divides the training process into two separate encoder-decoder networks: one for encoding actual attribute values and another for encoding differences from average values. Each network is trained independently on specific tasks, allowing unsupervised training while maintaining accuracy through the decomposition of the encoding task into manageable components.
Solution Approach 2:
The patent introduces an intermediary transformation layer that converts the output of the second network (difference encoding) into a format compatible with the first network's latent space. This intermediary mechanism enables the combination of both encoding approaches, allowing the system to leverage unsupervised training while achieving high accuracy through the mediation of the transformation layer.
2Measurement precision
If a dual network system with predefined structure is used, then unsupervised training accuracy is improved, but device complexity increases
Solution Approach 1:
The patent designs both encoder-decoder networks to operate on the same input data and produce outputs that can be combined through the intermediary transformation. The predefined data structure for latent space serves as a universal framework that both networks can utilize, allowing the system to handle multiple encoding tasks (actual values and differences) within a unified architectural paradigm, thereby managing complexity through reusability.
3Device complexity
If attribute values are encoded directly without comparison to averages, then encoding simplicity is maintained, but relative attribute information is lost
Solution Approach 1:
The patent segments the encoding process into two independent parts: one network encodes actual attribute values while the other encodes deviations from average values. This segmentation allows each network to focus on a specific aspect of the data, preserving both absolute and relative information without requiring complex interactions between encoding processes.
Data Source
AI summary
In some implementations at an electronic device, training a dual EDNN includes defining a data structure of attributes corresponding to defined parts of a task, processing a first instance of an input using a first EDNN to produce a first output while encoding a first set of the attributes in a first latent space, and processing a second instance of the input using a second EDNN to produce a second output while encoding attribute differences from attribute averages in a second latent space. The device then determines a second set of the attributes based on the attribute differences and the attribute averages. The device then adjusts parameters of the first and second EDNNs based on comparing the first instance of the input to the first output, the second instance of the input to the second output, and the first set of attributes to the second set of attributes.