Multi-view Embedding Softmax Compatibility Zero-Shot Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing zero-shot learning models, such as Embarrassingly Simple Zero-shot Learning (ESZSL), fail to effectively utilize information from multiple views of semantic representations and rely on naive optimization formulations, leading to limitations in pattern recognition systems.
Innovation Solution
A pattern recognition system that employs multi-view embedding by mapping low-level features and high-level semantic representations into a common space, using a soft-max formulation to enhance compatibility function values and suppress un-matched class representations, thereby improving inference accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If ESZSL uses a bilinear model to bridge low-level features and high-level semantic representation, then the optimization problem can be formulated, but it fails to utilize information from multiple views of semantic representation
Solution Approach 1:
The patent segments the semantic representation into multiple views (e.g., different semantic spaces or attribute sets) and processes each view separately through the bilinear model. This allows the system to utilize information from multiple views while maintaining the simplicity of the bilinear formulation for each individual view.
Solution Approach 2:
The patent extends the bilinear model to serve multiple functions by applying it across different semantic views simultaneously. The same basic bilinear framework is reused for each view, making the model multi-functional without requiring fundamentally different architectures for each semantic representation.
2Ease of manufacture
If ESZSL uses a best-fit optimization formulation with target vectors of +1 and -1, then the optimization problem is simple to formulate, but the bilinear model output range of (-∞, +∞) cannot naturally fit the discrete target values
Solution Approach 1:
The patent transforms the discrete target values (+1, -1) into continuous probability values through the soft-max function. This parameter transformation allows the bilinear model's continuous output range (-∞, +∞) to be naturally mapped to probability values (0, 1), improving fitting accuracy while maintaining optimization simplicity.
Solution Approach 2:
The patent replaces the direct best-fit approach with a probabilistic substitution using the soft-max function. Instead of directly fitting discrete target values, the system substitutes a probability distribution model that naturally handles the continuous-to-discrete mapping issue.
3Ease of operation
If existing zero-shot learning models treat compatibility function values independently for each class, then the computation is straightforward, but the system cannot provide confidence measures for classification outcomes
Solution Approach 1:
The patent introduces feedback through the soft-max function, which normalizes the compatibility scores across all classes to produce a probability distribution. This feedback mechanism provides confidence measures by showing not just which class has the highest compatibility, but also the relative confidence compared to other classes.
Solution Approach 2:
The patent applies soft-max normalization to create equipotential probability values across all class predictions. This ensures that the compatibility function values are transformed into a consistent probability scale, allowing for reliable confidence assessment while maintaining computational efficiency through vectorized operations.
4Loss of information
If the system maps low-level features and high-level semantic representations into a common space, then multi-view information can be integrated, but the dimensionality transformation increases computational complexity
Solution Approach 1:
The patent merges multiple semantic views into a common space by applying the same bilinear transformation framework to each view. This combining approach integrates multi-view information while avoiding the need for separate complex transformation systems for each view, thus reducing overall computational power requirements.
Data Source
AI summary
Described is a system for multi-view embedding for object recognition. During operation, the system receives an input image and transforms raw data of objects in the image with corresponding labels into low-level features and high-level semantic representations of the labels. A trained object recognition model is generated by embedding the low-level features with multiple high-level semantic representations. The system then receives data of an unknown object and assigns a label to the unknown object using the trained object recognition model. Finally, a device can be controlled based on the label.