Invariant Integration Layer for Data-Efficient Object Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training deep neural networks for object classification in mobile platforms is labor-intensive and time-consuming due to the need for extensive labeled data, which is difficult and costly.
Innovation Solution
Incorporating an invariant integration layer into convolutional neural networks that ensures geometric invariances such as rotation, translation, and scaling invariances, reducing the requirement for labeled data by leveraging geometric prior knowledge.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional convolutional neural networks are trained using large amounts of labeled data, then classification accuracy is improved, but the time and cost required for data labeling increases
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing geometric transformation matrices and invariant features during the training process. The invariant integration layer pre-processes feature maps by integrating over geometric transformations before classification, reducing the need for extensive labeled data and manual annotation time.
Solution Approach 2:
The patent changes the parameter representation by introducing invariant integration layers that transform feature maps into invariant representations. This parameter transformation allows the network to achieve rotation, translation, and scaling invariance, thereby reducing the amount of labeled data needed while maintaining classification accuracy.
2Measurement precision
If conventional convolutional neural networks are trained using large amounts of labeled data, then classification accuracy is improved, but the cost of data labeling increases
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing geometric transformation matrices and invariant features during the training process. The invariant integration layer pre-processes feature maps by integrating over geometric transformations before classification, reducing the need for extensive labeled data and manual annotation time.
Solution Approach 2:
The patent changes the parameter representation by introducing invariant integration layers that transform feature maps into invariant representations. This parameter transformation allows the network to achieve rotation, translation, and scaling invariance, thereby reducing the amount of labeled data needed while maintaining classification accuracy.
3Loss of time
If the amount of labeled data is reduced, then the time and cost for data labeling is reduced, but classification accuracy may deteriorate
Solution Approach 1:
The patent introduces an intermediary invariant integration layer between the convolutional layers and the classification layer. This intermediary layer integrates feature maps over geometric transformations, creating invariant representations that enable accurate classification with reduced labeled data by mediating between raw pixel data and classification decisions.
Solution Approach 2:
The patent changes the parameter representation by introducing invariant integration layers that transform feature maps into invariant representations. This parameter transformation allows the network to achieve rotation, translation, and scaling invariance, thereby reducing the amount of labeled data needed while maintaining classification accuracy.
4Quantity of substance
If geometric invariances are explicitly ensured through invariant integration layers, then data efficiency is improved, but the network architecture complexity increases
Solution Approach 1:
The patent applies universality by designing the invariant integration layer to handle multiple geometric transformations (rotation, translation, scaling) simultaneously within a single layer structure. This multi-functional layer achieves data efficiency improvements without proportionally increasing architecture complexity, as one layer handles multiple invariance requirements.
Data Source
AI summary
A method for generating a trained convolutional neural network including at least one invariant integration layer for classifying objects of a digital image of the surroundings of a mobile platform including a plurality of training cycles. Each training cycle includes: providing a digital image of the surroundings of a mobile platform including at least one object; providing a reference image associated with the digital image, the at least one object being labeled in the reference image; providing the digital image as an input signal of the convolutional neural network including at least one invariant integration layer; and adapting the convolutional neural network including at least one invariant integration layer in order to minimize a deviation of the classification from the particular associated reference image upon the classification of the at least one object of the digital image.