Prompt Learning for Suppressing Undesirable Image Classes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image generation systems face challenges in reducing the likelihood of outputting undesirable images in response to prompt inputs, particularly when using machine learning models for image generation, which can be costly and time-consuming to train.
Innovation Solution
An image acquisition device employs an image feature extraction and prompt feature extraction mechanism to calculate similarity between input images and combined prompts, updating control prompts using an evaluation function to minimize the likelihood of outputting images from suppression target classes, utilizing a learned machine learning model for efficient prompt learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning models are trained to reduce the likelihood of outputting undesirable images, then the quality and reliability of image output is improved, but the training cost and time consumption increase significantly
Solution Approach 1:
The patent applies preliminary action by pre-training the machine learning model on a large dataset of images with associated class labels before deployment. This pre-training phase prepares the model in advance to recognize and suppress undesirable image classes, so that during actual operation, the model can quickly filter unwanted images without requiring additional training time. The control prompt update mechanism also performs preliminary adjustments by updating prompts based on suppression target classes before image generation occurs.
Solution Approach 2:
The patent extracts the undesirable image classes by identifying specific suppression target classes from the overall image dataset. The control prompt update mechanism selectively extracts and updates prompts associated with these unwanted classes, separating them from the general image generation process. This extraction approach allows the model to focus on suppressing specific problematic classes without retraining on the entire dataset, thereby reducing training time while maintaining reliability.
2Reliability
If machine learning models are trained to reduce the likelihood of outputting undesirable images, then the reliability of image output is improved, but the computational resources and training cost increase
Solution Approach 1:
The patent extracts only the necessary components for suppression by identifying and updating control prompts specifically associated with suppression target classes. Instead of retraining the entire model, the system extracts and updates only the relevant prompt components that need adjustment, significantly reducing computational resources and training costs while maintaining improved reliability for suppressing undesirable images.
Solution Approach 2:
The patent applies parameter changes by updating the control prompt parameters rather than retraining the entire model. The control prompt update mechanism modifies specific parameters (prompts) associated with suppression target classes, changing the model's behavior for unwanted image classes without requiring full model retraining. This parameter-based approach reduces computational cost and energy consumption while achieving the desired reliability improvement.
3Measurement precision
If the control prompt is updated using an evaluation function to suppress specific classes, then the precision of image classification is improved, but the device complexity increases
Solution Approach 1:
The patent introduces an intermediary control prompt update mechanism that acts as a mediator between the image classification process and the prompt generation. This intermediary component calculates similarity between input images and reference images, then updates control prompts based on suppression target classes. By inserting this intermediary layer, the system achieves precise classification without requiring fundamental changes to the core model architecture, thus managing device complexity while improving measurement precision.
Solution Approach 2:
The patent implements feedback through the control prompt update mechanism, which uses evaluation functions to assess the similarity between images and adjusts control prompts accordingly. The system provides feedback by updating prompts based on suppression target class identification, creating a closed-loop system that continuously improves classification precision. This feedback mechanism achieves high measurement precision through iterative refinement rather than complex architectural changes.
Data Source
AI summary
An image acquisition device performs prompt learning using an evaluation function that indicates a worse evaluation the higher the similarity between an image feature vector, which is the feature vector of the input image, and the prompt feature vector, which is the feature vector of a combined prompt formed by combining a base prompt indicating a class in image classification and the input image class and a control prompt, which is data to be updated in a case where the class of an input image used to learn a prompt is a suppression target class, which is a class in which image output should be suppressed, and indicates a better evaluation the higher the similarity between the image feature vector and the prompt feature vector in a case where the input image class is a class other than the suppression target class, and acquires an image using the learned prompt.