Training Dynamics Prediction Model for Active Learning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing active machine learning methods face challenges in efficiently inferring the uncertainty of unlabeled data for effective model training, particularly due to high computational requirements and the need for labeled data to measure data difficulty, which limits their scalability and accuracy.

Innovation Solution

A training dynamics prediction model is introduced to quantify the uncertainty of unlabeled data by predicting training dynamics, allowing for selective data labeling and reducing computational burden, enabling efficient data selection and model training with reduced workforce and cost.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If training dynamics are used to infer uncertainty of unlabeled data, then data selection accuracy is improved, but computational cost increases excessively due to requiring probability calculation for each epoch

Engineering Contradiction:
Improveuncertainty inference accuracyVSAvoidcomputational cost
Core Design Contradiction:
Measurement precisionVSUse of energy by stationary object

Solution Approach 1:

The patent pre-calculates and stores training dynamics information (probability changes across epochs) for labeled data during the training phase. This pre-computed information is then used to infer uncertainty of unlabeled data without requiring full epoch-by-epoch probability calculations, thus reducing computational cost while maintaining inference accuracy.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent creates a simplified representation or copy of the training dynamics pattern from labeled data and applies it to estimate uncertainty of unlabeled data. Instead of performing complete training dynamics analysis on all data, it uses the learned pattern as a proxy, significantly reducing computation while preserving the essential uncertainty information.

Inventive Principle:
Principle #26Copying

2Productivity

If active machine learning is implemented to reduce human labeling workload, then productivity is improved, but measurement precision of data difficulty deteriorates due to lack of labeled data

Engineering Contradiction:
Improvelabeling efficiencyVSAvoiddata difficulty assessment
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent incorporates feedback from the training process by monitoring how probability distributions change across epochs for labeled data. This feedback signal (training dynamics) is then used to assess the difficulty of unlabeled data, creating a closed-loop system that continuously improves data selection based on actual training performance rather than static heuristics.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent changes the parameter used for uncertainty estimation from static metrics (like model confidence alone) to dynamic parameters that capture training progression (probability changes across multiple epochs). This parameter transformation enables more accurate difficulty assessment by reflecting how the model actually learns from different data points over time.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4242940A1Apparatus for active learning using training dynamics and method thereof
Publication Date: 2023.09.13 HYPERCONNECT INC
  • EP4242940A1 patent drawingFigure 1
  • EP4242940A1 patent drawingFigure 2
  • EP4242940A1 patent drawingFigure 3A

AI summary

A method for training a training dynamics prediction model (120) comprising acquiring classification information on training data included in a first dataset based on a classification model (110), acquiring target training dynamics information based on the classification information and a set of one or more classification information acquired based on the classification model (110) in one or more previous epochs, acquiring predictive training dynamics information on the training data based on the training dynamics prediction model (120), and training the training dynamics prediction model (120) based on the target training dynamics information and the predictive training dynamics information is disclosed.