Feature Tensor Distribution Distance for Open-Set Image Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks struggle with accurately classifying input data that are similar in pixel distribution and middle-level semantics but semantically different, particularly in open-set scenarios where new classes are recognized during production, and are prone to misclassifying rare examples due to limited training datasets.
Innovation Solution
A method involving a feed-forward neural network that processes input data through at least one layer to obtain a feature tensor, estimates a distribution of this tensor, determines a distance value between distributions, and uses this distance to improve classification by adjusting the neural network's output, allowing for better recognition of rare examples and out-of-distribution data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If cross-entropy loss function is used to train the neural network, then the model achieves good classification accuracy for common classes in the training dataset, but it fails to accurately classify rare examples and out-of-distribution data during production
Solution Approach 1:
The patent changes the parameter being optimized from simple classification accuracy to a distribution-based metric. By estimating the distribution of feature tensors and comparing it with training data distributions, the system detects out-of-distribution examples and rare cases, thereby improving reliability without sacrificing accuracy for common classes
Solution Approach 2:
The patent introduces an intermediary distribution estimation mechanism between the neural network output and the final classification decision. This intermediary layer estimates the distribution of feature tensors and compares it with reference distributions from training data, serving as a mediator to identify unreliable predictions
2Productivity
If the neural network is trained only on available training data, then the training process is efficient and completes quickly, but the model cannot recognize new classes or out-of-distribution data during production phase
Solution Approach 1:
The patent performs preliminary action by estimating the distribution of feature tensors during the training phase and storing reference distributions. This preliminary distribution estimation enables the model to later detect out-of-distribution data without requiring additional training, thus maintaining training efficiency while improving adaptability
Solution Approach 2:
The patent makes the neural network universally applicable to both seen and unseen classes by adding distribution-based detection capability. The same trained model can handle closed-set classification tasks while also detecting open-set out-of-distribution examples, achieving multi-functionality without retraining
3Device complexity
If the neural network uses standard feature extraction without distribution analysis, then the system complexity remains low and computation is fast, but the system cannot distinguish between semantically different data with similar pixel distributions
Solution Approach 1:
The patent adds another dimension to the analysis by moving from pixel-level or feature-level comparison to distribution-level comparison. By estimating and comparing the distributions of feature tensors, the system gains the ability to distinguish semantically different data with similar appearances without significantly increasing complexity
Data Source
AI summary
The present disclosure relates to processing input data by a neural network. Methods and apparatuses of some embodiments process the input data by at least one layer of the neural network and obtain thereby a feature tensor. Then, the distribution of the obtained feature tensors estimated. Another distribution is obtained. Such other distribution may be a distribution of another input data, or a distribution obtained by combining a plurality of distributions obtained for respective plurality of some input data. Then a distance value indicative of a distance between the two distributions is calculated and based thereon, a characteristic of the input data is determined. The characteristic may be pertinence to a certain class of data or a detection of out-of-distribution data or determination of reliability of a class determination or the like.