Resource-Efficient Teacher-Student Distillation with Style Augmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Training a student neural network with a smaller architecture to mimic the behavior of a larger teacher neural network is computationally expensive and resource-intensive, especially when the student network needs to generalize to unseen conditions not covered by the training data.

Innovation Solution

A method involving style augmentation of training images to expand the manifold of training data, using a teacher neural network to process both original and style-augmented versions, and optimizing student network parameters with a loss function to align outputs and intermediate work products, allowing for efficient training and generalization.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Use of energy by moving object

If a student neural network with smaller architecture is trained to mimic the teacher neural network, then resource consumption and computational cost are reduced, but the model's ability to generalize to unseen conditions deteriorates

Engineering Contradiction:
Improvecomputational costVSAvoidgeneralization ability
Core Design Contradiction:
Use of energy by moving objectVSAdaptability or versatility

Solution Approach 1:

The patent applies preliminary action by performing style augmentation on training images before training the student network. Multiple style-augmented versions of each training image are generated in advance, expanding the training manifold. This preliminary expansion of training data enables the smaller student network to learn more robust features that generalize better to unseen conditions, counteracting the typical generalization deterioration that would occur with model size reduction.

Inventive Principle:
Principle #10Preliminary action

2Quantity of substance

If the training dataset is expanded through style augmentation, then the manifold of training images is significantly expanded, but the complexity of the training process increases

Engineering Contradiction:
Improvenumber of training imagesVSAvoidtraining process complexity
Core Design Contradiction:
Quantity of substanceVSDevice complexity

Solution Approach 1:

The patent applies universality by using a single style augmentation module that can generate multiple style-augmented versions of training images. This modular approach allows the same augmentation process to be applied universally across the entire training dataset, efficiently expanding the number of training images without proportionally increasing training process complexity. The loss function also universally compares student and teacher network outputs across all augmented images, streamlining the overall training workflow.

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

Data Source

PatentEP4604012A1Transferring knowledge from a teacher neural network to a student neural network
Publication Date: 2025.08.20 ROBERT BOSCH GMBH
  • EP4604012A1 patent drawingFigure 1
  • EP4604012A1 patent drawingFigure 2a
  • EP4604012A1 patent drawingFigure 2b

AI summary

A method (100) for training a student neural network (2) to adopt the behavior of a given teacher neural network (1) that is trained to perform a given processing on input images, comprising the steps of: • providing (110) a set of training images x; • producing (120), from at least one training image x, one or more style-augmented versions x' that ∘ have the same semantic content as the original training image x but ∘ differ from this original training image x in their style; • processing (130) the training images x and the augmented versions x' by the teacher neural network (1) on the one hand, and by the student neural network (2) on the one hand; • evaluating (140), using a predetermined loss function (3), to which extent outputs (2b) and/or intermediate work products (2a) produced by the student neural network (2) from each image x, x' are in agreement with the outputs (1b) and/or intermediate work products (1a) produced by the teacher neural network (1) from the same image x, x'; and • optimizing (150) parameters (2c) that characterize the behavior of the student neural network (2) towards the goal of improving the value (3a) of the loss function (3).