Multi-Scale Neural Network Distillation for Dense Predictions

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing systems and methods for extracting relevant information from input data, such as images, face challenges in achieving high performance and efficiency, particularly for dense prediction tasks.

Innovation Solution

A data interpretation system utilizing a neural network with a backbone, task prediction units, feature propagation units, and distillation units to perform multiple tasks at different scales, enhancing performance through feature aggregation and propagation across scales.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If a neural network processes input data at a single scale, then the computational complexity is reduced, but the performance for dense prediction tasks deteriorates

Engineering Contradiction:
Improvecomputational complexityVSAvoidperformance for dense prediction tasks
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The network processes the input datum at multiple different scales simultaneously, dividing the processing into distinct scale-specific pathways. Each scale captures different levels of detail, with finer scales capturing local details and coarser scales capturing global context, thereby improving dense prediction performance without overwhelming computational complexity

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent adds the scale dimension to the processing architecture by implementing feature aggregation units that combine features from multiple scales. This multi-scale approach transforms the single-scale processing into a multi-dimensional feature space, enabling the network to capture both fine-grained and coarse-grained information for superior dense prediction performance

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Device complexity

If multiple tasks are performed simultaneously without task prediction units, then the system simplicity is maintained, but the performance for specific tasks deteriorates

Engineering Contradiction:
Improvesystem simplicityVSAvoidperformance for specific tasks
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent implements task prediction units that are specialized for specific tasks, with each unit receiving features at appropriate scales and performing task-specific processing. This allows different parts of the network to have different qualities optimized for their specific functions, improving overall task performance while maintaining reasonable system simplicity through modular design

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The backbone generates multi-scale features that serve multiple tasks simultaneously, and the feature aggregation units combine these features for different output tasks. This multi-functional design allows the same feature extraction infrastructure to support multiple tasks, improving performance across tasks while avoiding the need for completely separate processing pipelines for each task

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Productivity

If features are not aggregated across scales, then the processing speed is maintained, but the quality of dense predictions deteriorates

Engineering Contradiction:
Improveprocessing speedVSAvoidquality of dense predictions
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The network performs feature extraction at multiple scales in parallel during the forward pass, preparing multi-scale features beforehand. The feature aggregation units then combine these pre-computed features efficiently, avoiding the need for repeated processing and maintaining processing speed while improving prediction quality through comprehensive multi-scale feature utilization

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4091104B1Data interpretation system, vehicle, method for interpreting input data, computer program and computer-readable medium
Publication Date: 2025.09.17 TOYOTA JIDOSHA KK
  • EP4091104B1 patent drawingFigure 1
  • EP4091104B1 patent drawingFigure 2
  • EP4091104B1 patent drawingFigure 3

AI summary

A data interpretation system based on a neural network (1000) for performing output task(s) and which includes several units:. A backbone, based on said datum, calculates a backbone feature for several scales;. Task prediction units (TPUs) for each of said scales, based at least on the backbone feature (BFs) at said scale (Ss), calculate a task feature (Fi k,s) for each of said inside tasks (Tk);. A distillation unit (DU), based on the task features (Fi k,s) for each of said inside tasks and for each scale (Ss) among said scales, calculates an output feature (Fo k,s) for each of the at least one output task (T1,T2) for each of said scales; and. feature aggregation unit(s), for each of the at least one output task (T1,T2), combine(s) the output features (Fo k,s) calculated for the output task across the respective scales (Ss) so as to yield a prediction (DM,SI) for the output task (T1,T2).