Combined Learning Model Transfer for Multi-Task Signal Processing

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing multi-task deep neural networks (DNNs) face issues with useless calculation due to shared model parameters, leading to increased overall calculation amounts and difficulty in independently adjusting task performance.

Innovation Solution

A model generation device and method that involves transferring and combining transferrable learning models to generate a combined model using a transferable learning model, allowing for the generation of a combined model by transferring a part of the learning model to another transferrable learning model and learning the non-transfer portion, thereby generating a combined model.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If multiple independent DNNs are used for each task, then task performance can be independently adjusted, but useless calculation increases and overall calculation amount increases

Engineering Contradiction:

Improveindependent task performance adjustmentVSAvoiduseless calculation amount

Core Design Contradiction:

Adaptability or versatilityVSLoss of energy

Solution Approach 1:

The patent segments the DNN into task-specific portions and shared portions. Each task has its own specific layers that can be independently adjusted, while common layers are shared across tasks. This segmentation allows independent task performance adjustment without requiring completely separate DNNs, thereby reducing redundant calculations while maintaining adaptability.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent implements multi-functionality by creating a shared portion of the DNN that can be universally applied across multiple tasks. The shared layers process common features that are useful for multiple tasks, eliminating the need for each task to have its own complete DNN. This universal portion reduces overall calculation amount while still allowing each task to maintain its specific requirements through task-specific portions.

Inventive Principle:

Principle #6Universality (Multi-functionality)

2Loss of energy

If shared layers are used in multi-task DNN, then calculation efficiency improves, but task performance cannot be independently adjusted

Engineering Contradiction:

Improvecalculation efficiencyVSAvoidindependent task performance adjustment

Core Design Contradiction:

Loss of energyVSAdaptability or versatility

Solution Approach 1:

The patent divides the DNN architecture into distinct task-specific portions and shared portions. The task-specific portions can be independently adjusted for each task's performance requirements, while the shared portions provide calculation efficiency through parameter sharing. This segmentation resolves the contradiction by allowing both independent adjustment and efficient calculation to coexist in different parts of the model.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent applies local quality by giving different parts of the DNN different functions: the shared portion uses parameter sharing for calculation efficiency, while the task-specific portions use independent parameters for specialized task performance. Each portion has optimized quality suited to its specific purpose, allowing the overall system to achieve both efficiency and adaptability.

Inventive Principle:

Principle #3Local quality

3Adaptability or versatility

If completely separate DNNs are used for each task, then task independence is achieved, but overall calculation amount increases due to redundant calculations

Engineering Contradiction:

Improvetask independenceVSAvoidoverall calculation amount

Core Design Contradiction:

Adaptability or versatilityVSProductivity

Solution Approach 1:

The patent merges multiple task-specific DNNs into a single combined model by identifying and consolidating shared layers. The task-specific portions remain separate and independent, maintaining task independence, while the shared portions are merged to eliminate redundant calculations. This combining approach achieves both task independence and reduced overall calculation amount.

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The patent segments the overall model into independent task-specific portions and shared portions. The task-specific portions maintain independence for each task, while the shared portions are identified and merged across tasks. This segmentation strategy allows the system to keep task independence where needed while reducing redundant calculations through shared processing.

Inventive Principle:

Principle #1Segmentation

Data Source

PatentUS20250391401A1Model generation device, model generation method, signal processing device, signal processing method, and program

Publication Date: 2025.12.25 SONY GROUP CORP

AI summary

The present technology relates to a model generation device, a model generation method, a signal processing device, a signal processing method, and a program capable of suppressing useless calculation and independently adjusting performance of signal processing. A learning unit learns a transferable learning model, transfers a part of the learning model to another transferable learning model, and learns a non-transfer portion other than a transfer portion of the another learning model. A combination unit generates a combined model in which the non-transfer portion of the another learning model is combined with the learning model. The present technology can be applied to, for example, a case of generating a learning model that performs a plurality of pieces of signal processing.