Dual-Head Neural Network Training for Target Domain Accuracy

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing domain adaptation methods in machine learning fail to prioritize target labels over source labels, leading to reduced accuracy in transfer learning, especially when the target dataset is limited.

Innovation Solution

A neural network model is split into two portions, with one head for source data and one for target data, where the target head is prioritized by updating parameters based on consistency calculations and a pre-defined threshold, ensuring sufficient knowledge transfer from source to target.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If source and target labels are mixed equally during training, then the neural network learns from both domains, but the network cannot prioritize target label patterns over source label patterns

Engineering Contradiction:

Improvedomain adaptation capabilityVSAvoidtarget classification accuracy

Core Design Contradiction:

Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The neural network is segmented into multiple heads, with each head specialized for processing specific types of classification data (source domain or target domain). This segmentation allows the network to treat different domains separately rather than mixing them equally, enabling the target head to prioritize target label patterns while the source head handles source data independently.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

Different heads of the neural network are assigned different local qualities or specializations - the target head is optimized for target domain data while the source head handles source domain data. This local quality differentiation ensures that each head can focus on its specific domain without interference, allowing the target head to prioritize target label patterns when making classification decisions.

Inventive Principle:

Principle #3Local quality

2Quantity of substance

If more source labels are available, then the network has more training data, but accuracy on target domain suffers

Engineering Contradiction:

Improveamount of training dataVSAvoidtarget domain accuracy

Core Design Contradiction:

Quantity of substanceVSMeasurement precision

Solution Approach 1:

The network segments training data processing by directing source domain data to the source head and target domain data to the target head. This segmentation ensures that even when large amounts of source data are available, they are processed separately and do not interfere with the target head's ability to prioritize target label patterns, thus maintaining target domain accuracy.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

Instead of having the network prioritize source data when more source labels are available (the conventional approach), the invention inverts this logic by ensuring the target head maintains priority for target domain accuracy. The architecture allows the network to utilize abundant source data while still prioritizing target domain performance through the specialized target head.

Inventive Principle:

Principle #13The other way round (Inversion)

3Adaptability or versatility

If the neural network is split into multiple heads for different data types, then domain adaptation capability is improved, but the complexity of the network architecture increases

Engineering Contradiction:

Improvedomain adaptation capabilityVSAvoidnetwork architecture complexity

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The neural network is segmented into multiple heads, each handling specific data types. This segmentation improves domain adaptation capability by allowing specialized processing for source and target domains while maintaining a relatively simple overall architecture that can be implemented by extending standard neural network designs with additional head modules.

Inventive Principle:

Principle #1Segmentation

Data Source

PatentEP3885995B1Method for configuring a neural network model

Publication Date: 2026.03.11 AUMOVIO AUTONOMOUS MOBILITY GERMANY GMBH

Figure 1
Figure 2
Figure 3~4

AI summary

The invention relates to a computer-implemented method (100) for configuring a neural network model, wherein the method comprises the following steps: providing (102) a neural network model; splitting (104) the neural network model into a first portion and a second portion, the second portion comprising a first head for classifying a first type of classification data and a second head for classifying the second type of classification data; pre-processing (106), in a training phase, the second type of classification data in the first portion, processing (108) the pre-processed second type of classification data in the first and second heads and determining a first result of the processing of first type of classification data in the first head and a second result of the processing of first type of classification data in the second head; calculating (110) the consistency between the first result and the second result; and configuring (112) the neural network model by updating a value of at least one parameter of the second head based on the calculated consistency.