Contrastive Learning Training Using Associative Metadata

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing contrastive learning methods for developing similarity models face challenges in selecting relevant pairs for training, leading to suboptimal performance and longer training times, especially when relying on associative metadata for music and other domains with large datasets.

Innovation Solution

A contrastive learning training approach that uses associative metadata to select anchor-positive pairs through a probabilistic method, introducing randomness in pair selection to promote diversity and prevent overfitting, allowing for more accurate similarity models and faster training times.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If traditional contrastive learning methods are used to select training pairs, then the model can be trained, but the training time is longer and convergence is slower

Engineering Contradiction:
Improvetraining speedVSAvoidtraining time
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-grouping data samples according to their metadata attributes before training begins. This pre-organization of data into groups based on shared characteristics allows the model to quickly access relevant training pairs during training, eliminating the need for time-consuming pair selection during the training process and enabling faster convergence.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent segments the training data into multiple groups based on metadata attributes such as genre, artist, or other relevant characteristics. This segmentation allows the contrastive learning model to efficiently select positive pairs within the same group and negative pairs from different groups, significantly reducing the time required for pair selection and accelerating training.

Inventive Principle:
Principle #1Segmentation

2Reliability

If random pair selection is used to promote diversity, then overfitting is prevented, but the relevance of training pairs decreases

Engineering Contradiction:
Improvemodel accuracyVSAvoidtraining efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent applies local quality by selecting training pairs based on local metadata characteristics rather than purely random selection. Within each metadata group, the method strategically selects positive pairs that share specific local attributes while selecting negative pairs from different local groups, ensuring both relevance and diversity in training pairs, which improves model accuracy without sacrificing training efficiency.

Inventive Principle:
Principle #3Local quality

3Reliability

If all possible pairs are considered for training, then comprehensive coverage is achieved, but the complexity of pair selection increases

Engineering Contradiction:
Improvetraining completenessVSAvoidpair selection complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent reduces pair selection complexity by segmenting the data space into metadata-based groups. Instead of considering all possible pairs across the entire dataset, the method only considers pairs within and between specific groups, dramatically reducing the search space while maintaining comprehensive coverage of relevant training scenarios through the grouped structure.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies preliminary action by pre-organizing data into metadata groups before training. This pre-grouping eliminates the need for complex real-time pair selection during training, as the grouping structure is already established and can be efficiently queried, reducing computational complexity while ensuring comprehensive training coverage.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4443349A1Training methodology for contrastive learning based on associative metadata
Publication Date: 2024.10.09 UTOPIA MUSIC AG
  • EP4443349A1 patent drawingFigure 1
  • EP4443349A1 patent drawingFigure 2
  • EP4443349A1 patent drawingFigure 3A~3B

AI summary

According to an aspect, there is provided an apparatus for performing the following. The apparatus maintains, in a memory, machine-learning model(s) for calculating similarity between two datasets. The apparatus obtains a dataset comprising data samples annotated with associative metadata and performs, for each data subset of the dataset formed by the associative metadata: dividing the data subset or a part thereof randomly to anchor and candidate datasets, evaluating, using the machine-learning model(s), similarity between the anchor and candidate datasets and selecting, for each data sample in the anchor dataset, a data sample of the candidate dataset based on the evaluating. The apparatus trains at least one or more machine-learning models using anchor-candidate data pairs.