Ranked-List Loss Encoder for Measurement Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning methods for evaluating measurement data, such as image data, face challenges in efficiently mapping data samples to machine-analyzable representations due to slow convergence caused by trivial pairs or triplets in deep metric learning, limiting their effectiveness in applications like automated driving and classification tasks.
Innovation Solution
A method for training an encoder that maps data samples to machine-analyzable representations using a ranked-list loss function, which maintains similarity between similar samples and separates dissimilar samples through a similarity measure and a cost function optimized with trainable parameters, allowing for a more nuanced consideration of prior knowledge.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep metric learning with pairwise or triplet loss functions is used to learn embeddings, then semantic similarity information between data points can be captured, but convergence is slow due to a large proportion of trivial pairs or triplets
Solution Approach 1:
The patent changes the loss function parameter from traditional pairwise or triplet loss to a structured loss function that incorporates multiple examples and exploits structured information between them. This parameter change in the objective function transforms the optimization landscape to achieve faster convergence while maintaining semantic similarity capture capability.
Solution Approach 2:
The patent introduces an intermediary ranking mechanism that mediates between the embedding learning process and the loss computation. By introducing ranked list loss as an intermediary, the system can prioritize non-trivial pairs/triplets and eliminate trivial ones from the optimization process, thereby speeding up convergence without sacrificing semantic similarity accuracy.
2Ease of manufacture
If conventional contrastive learning with binary positive/negative sample classification is used, then training can be simplified, but prior knowledge about the degree of similarity between samples cannot be fully utilized
Solution Approach 1:
The patent segments the binary classification of positive/negative samples into a multi-level ranking structure. Instead of treating all positive samples equally, the method divides them into different rank levels based on their degree of similarity to the query sample. This segmentation allows the system to utilize prior knowledge about varying degrees of similarity while maintaining a structured training approach.
Solution Approach 2:
The patent adds a new dimension to the contrastive learning framework by introducing ranked lists of positive samples. Instead of the traditional binary dimension (positive/negative), the system creates an ordinal dimension with multiple rank levels. This dimensional expansion enables the incorporation of prior knowledge about similarity degrees without complicating the training procedure excessively.
Data Source
Figure 1a
Figure 1b
Figure 1c
AI summary
Method (100) for training an encoder fe(x) that maps data samples x of measurement data to machine-readable representations z, comprising the steps: • a set X of training samples x is provided (110), wherein, in the context of a given application, a relation is defined to what degree two samples x1 and x2 are similar to each other; • a function fθ(x) parameterized with trainable parameters θ is provided (120), which maps samples x to representations z; • a similarity measure h(x1, x2) is provided (140), which assigns samples x1 and x2 a similarity of representations fe(xi) and fθ(x2), and/or of processing products of these representations fe(xi) and fθ(x2); • at least one query sample q is drawn from the set X of training samples x (150);• For this query sample q, o a ranked set P of positive samples p from the set X that are similar to the query sample q, and o a set N of negative samples n from the set X that are no longer similar to the query sample q are determined (160); • at least the parameters θ are optimized (170) with the aim that the similarity measures h(q, p) are ordered according to the rank of the positive samples p ∈ P and are greater than h(q, n) for all n ∈ N.;