Active Learning Image Selection Using Uncertainty and Diversity Metrics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing active learning methods for computer vision tasks, particularly in medical imaging, face inefficiencies due to class imbalance caused by selecting data with high uncertainty measures, leading to skewed feature learning and reduced prediction performance and robustness.
Innovation Solution
A computer-implemented method for active learning that selects images for labeling based on both uncertainty and similarity metrics, ensuring a diverse and informative training set by reducing similarity and increasing uncertainty, using Bayesian deep learning and Monte Carlo dropout for uncertainty estimation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If active learning selects samples based solely on uncertainty measure, then data efficiency is improved, but class imbalance occurs leading to reduced prediction performance and robustness
Solution Approach 1:
The patent changes the selection criteria from a single parameter (uncertainty) to multiple parameters (uncertainty and diversity metrics). By incorporating diversity metrics such as feature space distance or clustering-based diversity, the system modifies the selection parameters to simultaneously consider both uncertainty and class representation, thereby preventing class imbalance while maintaining data efficiency.
Solution Approach 2:
The patent creates a composite selection strategy that combines uncertainty-based sampling with diversity-based sampling. This composite approach integrates multiple selection criteria into a unified framework, where samples are selected based on a combination of their uncertainty scores and their diversity contribution to the training set, thus resolving the contradiction between efficiency and reliability.
2Measurement precision
If multiple models are trained to estimate uncertainty, then uncertainty estimation accuracy is improved, but development speed decreases
Solution Approach 1:
The patent enables a single model to perform self-assessment of uncertainty through Monte Carlo dropout. Instead of requiring multiple separate models, the system activates dropout during inference on a single trained model, allowing it to generate uncertainty estimates for its own predictions. This self-service approach maintains accurate uncertainty estimation while significantly improving development speed.
Solution Approach 2:
The patent introduces dynamic behavior to a single static model by activating dropout layers during inference. This dynamic approach allows the model to generate multiple stochastic predictions on the same input, enabling uncertainty estimation without training multiple separate models. The dropout probability is dynamically controlled during the inference process to achieve reliable uncertainty measures.
Data Source
AI summary
A computer-implemented method of active learning for computer vision in digital images, comprising: inputting labelled image training examples into an artificial neural network in a training phase; training a computer vision model using the labelled training examples; carrying out a prediction task on each image of an unlabelled training set of unlabelled, unseen images using the model; calculating an uncertainty metric for the predictions in each image of the unlabelled training set; calculating a similarity metric for the unlabelled training set representing similarities between the images in the training set; selecting images from the unlabelled training set, in dependence upon both the similarity metric and the uncertainty metric of each image, to design a training set for labelling which tends to both lower the similarity between the selected images and increase the uncertainty of the selected images.