Multi-Prediction Model for Anomaly Detection in Multimedia Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine-learning methods for anomaly detection in multimedia data are inefficient due to the imbalance between normal and abnormal data classes, and they often fail to accurately detect anomalies because they either incorrectly reconstruct abnormal data or poorly predict missing information.
Innovation Solution
A multi-prediction model is trained to reconstruct masked normal multimedia data using a set of predictors, each specializing in one possible prediction, which improves anomaly detection by not generalizing predictions well in the presence of anomalies and adapting better to normal data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised machine-learning methods are used for anomaly detection, then the detector can be trained on representative data of categories to be detected, but the task of annotating abnormal data becomes extremely expensive and requires covering a vast range of disparate anomalies
Solution Approach 1:
The patent inverts the traditional supervised learning approach by using unsupervised learning where the model is trained only on normal data without any anomaly annotations. The anomaly detection is achieved by measuring how well the model can reconstruct or predict the input data, with poor reconstruction indicating anomalies. This eliminates the need for expensive anomaly annotation while maintaining detection capability.
Solution Approach 2:
The model performs self-service by automatically learning the characteristics of normal data and using this knowledge to identify anomalies without requiring external annotation of abnormal cases. The system serves itself by using the training data to create a reference model that inherently detects deviations from normal patterns.
2Ease of manufacture
If reconstruction-based unsupervised methods are used, then the model can be trained on normal data only, but the model may incorrectly reconstruct abnormal data leading to non-detection
Solution Approach 1:
The patent segments the anomaly detection task into multiple specialized predictors, each trained to predict specific aspects or features of normal data. By dividing the reconstruction task among multiple specialists rather than using a single general-purpose predictor, the system achieves more reliable anomaly detection while maintaining ease of training on normal data only.
Solution Approach 2:
The patent changes the parameters and architecture of the prediction model to better capture the characteristics of normal data. By adjusting model parameters and using ensemble methods with multiple predictors, the system improves its ability to distinguish normal from abnormal patterns, enhancing detection reliability without requiring annotated anomaly data.
3Loss of information
If prediction-based unsupervised methods are used, then the model can learn to predict missing information, but the model often fails to accurately predict missing information leading to poor anomaly detection
Solution Approach 1:
The patent divides the prediction task into multiple specialized predictors, each focusing on predicting specific types of missing information or data aspects. This segmentation allows each predictor to specialize in certain patterns, improving overall prediction accuracy and anomaly detection capability compared to a single general-purpose predictor.
Solution Approach 2:
The patent combines the outputs of multiple specialized predictors to form a comprehensive prediction. By merging the strengths of different predictors that specialize in different aspects of data prediction, the system achieves superior accuracy in recovering missing information and detecting anomalies.
Data Source
AI summary
A method for training a model for reconstructing multimedia data represented by at least one modality, the model being composed of a set of a plurality of different predictors for each datum modality, the training method comprising the steps of, for each datum of a training dataset containing no anomalies, for each datum modality: masking at least part of the datum modality, training each predictor of the set associated with the modality to compute a different prediction of the same masked datum, each predictor being specialized in one possible prediction of the masked datum among various credible alternatives, selecting the predictor of the set that provides the closest prediction to a reference datum extracted from the training data, computing a distance between the prediction and the reference datum, computing a first cost function equal to the sum of the distances for all the modalities, updating the parameters of the predictors selected for each modality so as to minimize the first cost function.