Dataset Updating via Soft Label Similarity for Edge Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for updating datasets with limited sample data face challenges such as overfitting due to bias distribution, making it difficult to learn effectively from a small number of samples.
Innovation Solution
A method that involves acquiring a reference dataset with soft labels indicating the probability of each image belonging to target classes, determining the soft label for input images based on similarity with reference images, and adding these input images with their soft labels to the reference dataset for updating.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a machine learning model is trained using only a small amount of sample data, then the model can be updated in edge computing scenarios, but the bias distribution formed by the model is prone to overfitting and learning effectiveness deteriorates
Solution Approach 1:
The patent introduces a reference dataset with soft labels as an intermediary between the limited input data and the machine learning model. The reference dataset serves as a mediator that provides additional information about target classes, enabling the model to learn from limited samples without overfitting. The soft labels in the reference dataset act as a bridge, providing probability distributions that guide the model's learning process.
Solution Approach 2:
The patent performs preliminary action by pre-processing reference images to generate soft labels before they are used for model training. The reference dataset is prepared in advance with probability distributions for each target class, so that when new input data arrives at the edge device, the model can immediately utilize the pre-computed soft labels to enhance learning effectiveness without requiring complex real-time processing.
2Reliability
If complex model and loss function design is used to handle limited sample data, then learning effectiveness may improve, but computing costs increase
Solution Approach 1:
The patent uses copying by creating a reference dataset that replicates the structure and information of the original training data but with enhanced soft label information. Instead of designing complex models to handle limited data, the approach copies and augments the reference data with probability distribution labels, providing the model with additional learning signals without requiring complex computational architectures.
Solution Approach 2:
The patent applies parameter changes by transforming the data representation from hard labels to soft labels, which are probability distributions over target classes. This change in the label parameter allows the model to learn from limited samples more effectively by utilizing the probability information, thereby improving learning effectiveness without increasing model complexity or computing costs.
Data Source
AI summary
Embodiments of the present disclosure relate to a method, an electronic device, and a computer program product for dataset updating. The method includes: acquiring a reference dataset comprising a plurality of reference images, wherein each reference image has a soft label, and the soft label indicates a probability that the each reference image belongs to each target class in a plurality of target classes. The method may further include: determining, according to a similarity between an input image received and the each reference image, an image soft label for the input image. The method further includes: adding the input image and the image soft label for the input image to the reference dataset to update the reference dataset.