Latent Space Training Data Synthesis for Lightweight Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Lightweight deep learning faces challenges in reusing original training data due to privacy issues and the difficulty in optimizing high-dimensional raw input spaces, leading to low fidelity and diversity in generated data.
Innovation Solution
An apparatus and method that generate training data by using a combination of generators and convolution blocks to produce feature maps and feature enhancement maps, adjusting channel distributions with scaling parameters to match original training data, and iteratively refining images to achieve fidelity and diversity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If optimization of high dimension raw input space is performed to synthesize training data, then training data can be generated without access to original data, but the generated data has low fidelity to original data distribution and low diversity
Solution Approach 1:
The patent transforms the optimization problem from the high-dimensional raw input space to a low-dimensional latent space by introducing a latent variable z. This dimensional transformation enables effective optimization while preserving the fidelity to original data distribution through the encoder-decoder architecture that maps between latent space and image space.
Solution Approach 2:
The patent introduces an intermediary latent variable z that mediates between the optimization process and the final generated image. The latent variable serves as a bridge that allows optimization in a simplified space while maintaining the connection to the original data distribution through the encoder and decoder components.
2Productivity
If optimization of high dimension raw input space is performed, then training data synthesis is enabled, but mode collapse occurs and diversity is reduced
Solution Approach 1:
By moving the optimization to latent space rather than raw input space, the system can generate diverse images corresponding to different classes without mode collapse. The latent variable z can be sampled from a standard normal distribution, enabling the generation of varied images that maintain diversity while still synthesizing training data effectively.
3Device complexity
If averaging of image characteristic statistics is performed, then high dimension raw input space optimization is simplified, but specific information of individual images is lost
Solution Approach 1:
The latent variable z acts as an intermediary that captures the essential information of individual images without requiring direct optimization of the high-dimensional image space. The encoder maps individual image characteristics to the latent space, preserving specific information while simplifying the optimization process through the lower-dimensional latent representation.
Data Source
AI summary
An apparatus and method for generating training data according to an embodiment is disclosed. The apparatus for generating training data according to an embodiment includes at least one processor and a memory to store instructions for executing the at least one processor, wherein upon being executed by the at least one processor, the instructions allow the at least one processor to output a first image for one sample vector from a first generator included in the apparatus, and generate a second image from a second generator included in the apparatus based on the first image and a feature map extracted from a convolution block for each stage of a lightweight target model for the first image.