Unsupervised Classifier Training via Neighbor Likelihood Clustering

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for unsupervised learning of classifiers for image classification tasks are either sensitive to initialization conditions, prone to degenerate solutions, or require annotated data for fine-tuning.

Innovation Solution

A computer-implemented training method that learns a parametric model for classifying samples in a fully or largely unsupervised mode, using a pretext model trained in a self-supervised manner to learn semantically meaningful features, and a classifier trained to maximize the likelihood of samples and their neighbors belonging to the same cluster.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of information

If representation learning methods use self-supervised learning to generate feature representations, then the need for ground truth annotations is reduced, but a second step requiring annotated data for fine-tuning is still necessary

Engineering Contradiction:
Improveground truth annotationsVSAvoidtraining process complexity
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

The patent combines feature learning and clustering into a single end-to-end training process, eliminating the need for separate representation learning and fine-tuning steps. The classifier is trained directly on the dataset using a loss function that maximizes the likelihood of samples and their k-nearest neighbors belonging to the same cluster, achieving unsupervised classification without requiring annotated data for any training stage.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The method enables the system to perform unsupervised classification entirely on its own without external annotated data. The training process automatically learns meaningful cluster structures by exploiting the intrinsic relationships between samples and their neighbors, making the system self-sufficient for unsupervised learning tasks.

Inventive Principle:
Principle #25Self-service

2Loss of information

If offline clustering techniques like K-means are applied to learned feature representations, then annotated datasets are not required, but cluster degeneracy occurs and semantic meaningfulness is not guaranteed

Engineering Contradiction:
Improveground truth annotationsVSAvoidcluster quality
Core Design Contradiction:
Loss of informationVSReliability

Solution Approach 1:

The patent performs preliminary feature learning within the end-to-end training process itself, rather than relying on separate offline clustering. The classifier learns discriminative features that are optimized for the specific classification task, ensuring that the resulting clusters are semantically meaningful and reliable.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The training process incorporates feedback through the loss function that uses k-nearest neighbor relationships. The loss function continuously adjusts the classifier parameters based on the consistency between predicted cluster assignments and the intrinsic neighbor structure of the data, ensuring high-quality, non-degenerate clusters.

Inventive Principle:
Principle #23Feedback

3Device complexity

If end-to-end learning pipelines combine feature learning with clustering, then the process is integrated, but the methods are sensitive to initialization conditions and prone to converge towards degenerate solutions

Engineering Contradiction:
Improvetraining process integrationVSAvoidconvergence stability
Core Design Contradiction:
Device complexityVSReliability

Solution Approach 1:

The patent employs a specific loss function formulation that changes the optimization landscape to avoid degenerate solutions. The loss function incorporates k-nearest neighbor relationships and uses a temperature parameter to control the softness of cluster assignments, making the training process more stable and less sensitive to initialization conditions.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The training process dynamically adjusts the softness of cluster assignments through the temperature parameter, allowing the model to transition from exploratory phase to convergence phase. This dynamic behavior helps the system escape poor local minima and achieve stable, meaningful cluster solutions.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3913544B1A computer-implemented training method, classification method and system, computer program and computer-readable recording medium
Publication Date: 2025.03.12 TOYOTA JIDOSHA KK
  • EP3913544B1 patent drawingFigure 1
  • EP3913544B1 patent drawingFigure 2
  • EP3913544B1 patent drawingFigure 3

AI summary

A computer-implemented method for training a classifier (Φη), comprising: S10) training a pretext model (ΦΘ) to learn a pretext task, so as to minimize a distance between an output of a source sample via the pretext model (ΦΘ) and an output of a corresponding transformed sample via the pretext model (ΦΘ), the transformed sample being a sample obtained by applying a transformation (T) to the source sample ; S20) determining a neighborhood (NXi) of samples (Xi) of a dataset (SD) in the embedding space; S30) training the classifier (Φη) to predict respective estimated probabilities Φηj(Xi), j=1..C, for a sample (Xi) to belong to respective clusters (Cj), by using a second training criterion which tends to: - maximize a likelihood for a sample and its neighbors (Xj) of its neighborhood (Nxi) to belong to the same cluster; and - force the samples to be distributed over several clusters.