Synthetic Data Generation Using Primitive Features Against Overfitting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for machine perception, such as those used in autonomous systems, often suffer from overfitting due to a lack of diverse training data, leading to poor performance on unseen data, especially when generating photorealistic images.
Innovation Solution
A generative machine learning model is trained using a combination of real and other primitive features extracted from images, emphasizing important features to reduce overfitting and enhance the model's accuracy in producing photorealistic outputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a generative machine learning model is trained using only standard image data without emphasizing primitive features, then the training process is simpler and faster, but the model suffers from overfitting and produces less accurate photorealistic outputs
Solution Approach 1:
The patent applies local quality by emphasizing specific primitive features (edges, textures, colors, shapes) within the training data rather than treating all image data uniformly. The system extracts and prioritizes these fundamental visual elements to improve photorealism in specific aspects of the generated images, thereby increasing manufacturing precision of the synthetic training data without requiring complete redesign of the training process
Solution Approach 2:
The patent implements preliminary action by pre-extracting primitive features from real images before feeding them to the generative model. This preprocessing step prepares the training data in advance with emphasized primitive features, allowing the model to learn from structured information that reduces overfitting and improves accuracy without adding complexity during the actual training execution
2Reliability
If diverse real images are collected and used for training to prevent overfitting, then the model generalizes better to unseen data, but the cost and time required to capture and process such data increases significantly
Solution Approach 1:
The patent applies copying by using a generative machine learning model to create synthetic training data that replicates the characteristics of diverse real images. Instead of physically collecting and processing大量 real images, the system learns from a smaller set of real images and generates synthetic copies that maintain the essential primitive features and variability, thereby improving reliability without incurring the time cost of extensive data collection
Solution Approach 2:
The patent implements parameter changes by transforming real images into synthetic images through modifications of various parameters while preserving primitive features. The system varies parameters such as lighting, texture, color, and geometric transformations to generate diverse training samples from limited real data, achieving better generalization performance without the time investment required to capture equivalent diverse real images
3Manufacturing precision
If the generative model is trained to produce highly photorealistic images, then the synthetic training data quality improves, but the computational resources and training time required increase
Solution Approach 1:
The patent applies local quality by focusing computational efforts on emphasizing specific primitive features (edges, textures, colors, shapes) rather than attempting to optimize all aspects of image generation uniformly. This targeted approach improves photorealism in critical visual elements while reducing unnecessary computational energy consumption in less important areas
Solution Approach 2:
The patent implements preliminary action by pre-processing real images to extract and emphasize primitive features before training the generative model. This preparation step reduces the computational burden during actual training by providing structured input data, allowing the model to achieve high photorealism with reduced energy consumption compared to training on raw unprocessed images
Data Source
AI summary
A system for training a generative machine learning model, comprising: a hardware processor, configured for: extracting a plurality of real primitive features from a plurality of real images captured by a sensor in a physical environment; extracting a plurality of other primitive features from a plurality of other images depicting another physical environment; training a generative machine learning model to produce a refined image in response to a synthetic image, the training using the plurality of real images, the plurality of real primitive features, the plurality of other images and the plurality of other primitive features, where the generative machine learning model is adapted for receiving the plurality of real primitive features and additionally or alternatively the plurality of other primitive features as input, to produce a trained model; and providing the trained model to another hardware processor for the purpose of generating synthetic training data.