Open-Set Classification via Robust Loss Functions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing multi-class recognition systems are not robust to unknown inputs, as they fail to effectively separate features from known classes from those of unknown classes, leading to overlapping representations and high error rates when encountering novel data.
Innovation Solution
A classification system is developed that uses explicit representations for unknown classes, trained with a mixture of known and unknown data using robust loss functions like Entropic Open-Set Loss and Objectosphere Loss, which separates known inputs from unknown inputs by minimizing feature magnitude and maximizing entropy, driving unknown samples to a desired representation such as the origin.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional multi-class recognition systems are used, then known classes can be classified, but the system fails to separate unknown classes leading to overlapping representations and high error rates
Solution Approach 1:
The patent segments the classification problem into two distinct parts: known class classification and unknown class detection. By introducing a separate rejection mechanism that operates independently from the standard classification head, the system can handle known and unknown inputs through different processing paths, resolving the contradiction between maintaining high accuracy for known classes while achieving robustness against unknown inputs.
Solution Approach 2:
The patent introduces an intermediary rejection layer between the feature extractor and the final classification output. This intermediary component evaluates whether inputs belong to known or unknown classes before final classification, allowing the system to maintain high reliability for known classes while being adaptable to reject unknown inputs, thus resolving the technical contradiction.
2Measurement precision
If the system is trained only on known classes, then classification performance on known data is optimized, but the system cannot handle novel data from unknown classes
Solution Approach 1:
The patent applies preliminary action by training the rejection mechanism on a mixture of known and unknown data during the training phase. This preliminary exposure to both known and unknown distributions allows the system to learn discriminative features for rejection before deployment, enabling it to maintain high classification precision on known data while being adapted to handle novel unknown data during testing.
Solution Approach 2:
The patent changes the training parameters by using a mixed dataset containing both known and unknown samples during training, rather than only known classes. This parameter change in the training distribution enables the model to learn both classification and rejection capabilities simultaneously, achieving both measurement precision for known classes and adaptability for novel data.
3Productivity
If standard loss functions are used during training, then known class classification is optimized, but feature representations for unknown classes overlap with known classes
Solution Approach 1:
The patent merges multiple loss functions into a unified training objective: the standard classification loss for known classes and a rejection loss for unknown classes. By combining these losses during training, the system simultaneously optimizes both classification accuracy and feature separation quality, resolving the contradiction between training efficiency and feature separation precision.
Solution Approach 2:
The patent creates a composite loss function that combines different loss components (classification loss and rejection loss) with appropriate weighting. This composite loss function enables the training process to achieve both high productivity through efficient gradient-based optimization and high manufacturing precision through improved feature separation between known and unknown classes.
Data Source
AI summary
The invention includes systems and methods, including computer programs encoded on computer storage media, for classifying inputs as belonging to a known or unknown class as well as for updating the system to improve is performance. In one system, there is a desired feature representation for unknown inputs, e.g., a zero vector, and the system includes transforming input data to produce a feature representation, using that to compute dissimilarity with the desired feature representation for unknown inputs and combining dissimilarity with other transformations of the feature representation to determine if the input is from a specific known class or if it is unknown. In one embodiment, the system transforms the magnitude of the feature representation into a confidence score. In an update method to improve performance, the system transforms inputs into feature representations which go through a scoring means and then use a robust loss function, which has different loss terms for known and unknown inputs which are then used to update the system weights to improve performance.