Model Training via Intermediate Representation Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing adversarial training methods enhance model robustness to adversarial examples at the cost of reducing accuracy for clean samples.
Innovation Solution
A learning device that calculates deviations between outputs and intermediate representations of a teacher and learning models, updating the learning model's parameters to minimize these deviations and optimize both output and intermediate representation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If adversarial training is applied to enhance model robustness to adversarial examples, then robustness is improved, but accuracy for clean samples decreases
Solution Approach 1:
The patent segments the training objective into two distinct components: (1) minimizing output deviation between clean and adversarial examples, and (2) minimizing intermediate representation deviation between the student model and teacher model. This segmentation allows independent optimization of robustness and accuracy, resolving the contradiction by treating them as separate but complementary goals rather than conflicting objectives.
Solution Approach 2:
The patent introduces intermediate representations (feature embeddings from hidden layers) as an intermediary mechanism to transfer knowledge from the teacher model to the student model. This intermediary allows the student model to learn robust feature extraction patterns without directly copying final classifications, enabling robustness transfer while preserving clean sample accuracy through the mediating role of intermediate features.
2Reliability
If the model learns to classify adversarial examples as having the same label as clean samples, then robustness is improved, but the model may fail to extract features for clean sample classification
Solution Approach 1:
The patent moves the training objective from the final classification output dimension to the intermediate feature representation dimension. By minimizing deviation in intermediate representations (which exist in a high-dimensional feature space) rather than just final labels, the model learns robust feature extraction patterns in this intermediate dimensional space, preserving feature discrimination capability while achieving robustness.
Solution Approach 2:
The patent replaces the traditional adversarial training mechanism (which directly optimizes classification outputs) with a knowledge distillation mechanism that optimizes intermediate feature representations. This substitution allows the model to learn from the teacher model's robust feature extraction patterns without the mechanical constraint of directly matching final labels, preserving feature extraction capability while achieving robustness.
Data Source
AI summary
A learning device includes processing circuitry configured to calculate a degree of deviation between a first output obtained by inputting first training data to a learned first model and a second output obtained by inputting second training data created by giving noise to the first training data to a second model, and a degree of deviation between an intermediate representation of the first model generated in a process of obtaining the first output and an intermediate representation of the second model generated in a process of obtaining the second output, and update a parameter of the second model so that the degree of deviation between the first output and the second output and the degree of deviation between the intermediate representation of the first model and the intermediate representation of the second model are reduced.