Automated Dataset Generation via Iterative Discriminator Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for generating large, accurately labeled datasets for machine learning are labor-intensive, prone to inconsistent labeling due to human bias, and inefficient in producing diverse dataset types, particularly for proprietary or sensitive data, and often require significant computational resources.
Innovation Solution
An automated system that uses an analytic server to iteratively refine dataset generator models through pre-trained discriminators, enabling the creation of large, labeled datasets from a small set of seed examples using unsupervised or weakly supervised learning, reducing the need for extensive human effort and computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling methods are used to generate training datasets, then labeling accuracy can be maintained through skilled human effort, but the productivity and time required to produce large datasets deteriorate significantly
Solution Approach 1:
The system enables datasets to label themselves through self-supervised learning mechanisms. The model automatically generates labels by predicting missing data, inferring relationships, or completing patterns without human intervention, allowing the system to serve its own labeling needs at scale
Solution Approach 2:
Manual mechanical labeling processes are replaced with automated machine learning models that perform labeling tasks algorithmically. The system substitutes human cognitive processes with computational algorithms that can process and label data at machine speed while maintaining consistency
2Productivity
If crowd-sourcing is used to distribute labeling work among many people, then productivity increases, but labeling consistency deteriorates due to personal biases
Solution Approach 1:
The system ensures homogeneous labeling by using a single unified model architecture and consistent learning algorithms across all data processing. All data points are labeled through the same computational process, eliminating the heterogeneity introduced by different human labelers with varying biases and standards
3Manufacturing precision
If CGI technologies are used to generate photorealistic imagery, then the quality and realism of generated datasets improve, but the computational resources and skilled labor required deteriorate tremendously
Solution Approach 1:
The system performs preliminary actions by pre-training models on available data and pre-processing input data to reduce complexity before generation. By preparing models and data in advance with appropriate feature extraction and dimensionality reduction, the actual generation process requires significantly less computational power during deployment
4Adaptability or versatility
If GANs are used to generate datasets with multiple labels, then the versatility of the generated datasets improves, but the requirement for multiple separate training datasets and models deteriorates
Solution Approach 1:
The system creates a universal model architecture that performs multiple labeling functions simultaneously. A single model can generate datasets with various types of labels including object detection, segmentation, classification, and relational labels, eliminating the need for separate specialized models for each label type
Data Source
AI summary
Disclosed herein are embodiments of systems, methods, and products comprising an analytic server that automates training dataset generation for different application areas. The server may perform an automated, iterative refinement process to build a collection of dataset generator models over time. The server may receive a set of seed examples in a domain and generate candidate examples based on the features of the seed examples using data synthesis techniques. The server may execute a pre-trained label discriminator (LD) and domain discriminator (D2) on the candidate examples. The LD may identify and reject mislabeled data. The D2 may identify and reject out of domain data. The analytic server may regenerate new labeled data based on the feedback of the LD and D2. The analytic server may train a dataset generator by iteratively performing these steps for refinement until the regenerated candidate examples reach a pass rate threshold.