Slot-Based Model Training via Feature Map Decomposition

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current machine learning models for image processing tasks face inefficiencies in identifying and processing object instances, as they often require iterative and random initialization, leading to increased computation time, potential lock-in to local optima, and reduced reproducibility.

Innovation Solution

A method for training machine learning models that decomposes feature maps into components and initializes slots based on these components, allowing for deterministic and efficient assignment of features to slots, thereby reducing computation time and increasing the likelihood of reaching the global optimum.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If random initialization is used for slot assignment, then the model can explore the feature space broadly, but the computation time increases and the model may get locked in local optima

Engineering Contradiction:
Improveoptimization convergence to global optimumVSAvoidcomputation time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent applies preliminary action by performing feature map decomposition into N components before slot initialization. This preprocessing step creates an informed starting point for slot assignment, allowing the optimization to begin closer to the global optimum rather than from random initialization, thus reducing computation time while improving convergence reliability.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent changes the initialization parameter from random values to decomposed feature components. By transforming the initialization strategy from uniform random distribution to structured feature-based assignment, the model achieves better convergence properties and reduced computation time without sacrificing exploration capability.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If random initialization is used for slot assignment, then the model can start training, but the reproducibility of results decreases

Engineering Contradiction:
Improvereproducibility of training resultsVSAvoidtraining process complexity
Core Design Contradiction:
ReliabilityVSEase of operation

Solution Approach 1:

The decomposition of feature maps into N components is performed as a preliminary deterministic step before slot initialization. This creates a reproducible starting point that eliminates randomness from the initialization process, ensuring that training results can be reproduced across different runs while maintaining training simplicity.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If iterative training from random initialization is used, then the model can learn feature assignments, but the number of iterations required increases

Engineering Contradiction:
Improvefeature assignment accuracyVSAvoidtraining speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent performs feature map decomposition as a preliminary action that pre-organizes features into N components before the iterative training begins. This preprocessing reduces the distance to the optimal solution, allowing the iterative optimization to converge faster while maintaining the same level of feature assignment accuracy.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

By initializing slots based on decomposed feature components rather than random values, the patent effectively skips through the early iterations that would otherwise be needed to escape random initialization and approach the optimal assignment, thus accelerating training speed without compromising accuracy.

Inventive Principle:
Principle #21Skipping (Rushing through)

Data Source

PatentEP4343627A1Improved training of slot-based machine learning models
Publication Date: 2024.03.27 ROBERT BOSCH GMBH
  • EP4343627A1 patent drawingFigure 1
  • EP4343627A1 patent drawingFigure 2
  • EP4343627A1 patent drawingFigure 3

AI summary

A method (100) for training a machine learning model (1) that is configured to process measurement data (2) with respect to a given task, wherein said model (1) comprises a slot building part (11) that is configured to assign features (3) of an inputted record of measurement data (2) to a plurality of slots (4a-4d) and a task part (12) that is configured to process the features (3) of the record of measurement data (2) that are assigned to each slot (4a-4d) separately, the method (100) comprising the steps of: • providing (110) training samples (2a) for records of measurement data (2); • producing (120), from each training sample (2a), by at least one feature extraction layer (11a) in the slot building part (11), at least one feature map (3#); • decomposing (130) each feature map (3#) into N components (3a-3f) according to a predetermined decomposing criterion and/or method; • initializing (140) a predetermined number K of slots (4a-4d) by assigning, to these slots (4a-4d), features (3) from the feature map (3#), based at least in parts on the N components (3a-3f); and • training (150) the slot building part (11) and the task part (12) towards at least one given objective.