Feature Sub-Network Trainer for DNN Interpretability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks (DNNs) face challenges in interpretability due to their complex, non-linear mathematical space, leading to the 'black box' problem, where it is difficult to understand how they arrive at predictions, and the risk of converging to local minima rather than the global minimum, affecting trust and deployment, especially in regulated industries.
Innovation Solution
A feature sub-network trainer splits a pre-trained DNN into feature and dense sub-networks, with additional training focusing on sparse connected layers, reducing the dimensionality and likelihood of converging to local minima, thereby improving robustness and interpretability by forming a feature-trained DNN that is more likely to have fully converged to the global minimum.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If additional hidden layers are added to increase abstraction capability, then learning complexity improves, but computational resources and training time increase significantly
Solution Approach 1:
The patent segments the DNN into feature sub-network and dense sub-network, training the feature sub-network separately with fewer parameters. This segmentation allows the system to achieve high learning capability through the dense sub-network while reducing computational resources by pre-training and reusing the feature sub-network, thus resolving the contradiction between learning capability and computational resource consumption.
2Measurement precision
If DNN parameters increase to improve prediction accuracy, then learning performance improves, but interpretability deteriorates due to black box problem
Solution Approach 1:
The patent extracts the feature sub-network from the full DNN and trains it separately with fewer parameters. By taking out this critical component, the system maintains prediction accuracy through the dense sub-network while improving interpretability by analyzing the smaller, isolated feature sub-network, thus resolving the contradiction between prediction accuracy and interpretability.
3Reliability
If training DNN multiple times with different initial values to avoid local minima, then robustness improves, but training time and computational cost increase
Solution Approach 1:
The patent performs preliminary action by pre-training the feature sub-network once and reusing it across multiple training runs. This pre-trained feature sub-network serves as a robust foundation that reduces the need for multiple training runs with different initial values, thereby improving robustness while reducing the time and computational cost associated with repeated training.
4Speed
If DNN converges to local minima instead of global minimum, then training speed improves, but prediction reliability deteriorates
Solution Approach 1:
The patent segments the DNN into feature and dense sub-networks, training the feature sub-network separately. This segmentation creates a simpler optimization problem with fewer local minima, allowing the feature sub-network to converge reliably to the global minimum faster. The dense sub-network then builds upon this reliable foundation, maintaining both training speed and prediction reliability.
Data Source
AI summary
A feature sub-network trainer improves robustness of interpretability of a deep neural network (DNN) by increasing the likelihood that the DNN will converge to a global minimum of a cost function of the DNN. After determining a plurality of correctly classified examples of a pre-trained DNN, the trainer extracts from the pre-trained DNN a feature sub-network that includes an input layer of the DNN and one or more subsequent sparsely-connected layers of the DNN. The trainer averages output signals from the sub-network to form an average representation of each class identifiable by the DNN. The trainer relabels each correctly classified example with the appropriate average representation, and then trains the feature sub-network with the relabeled examples. In one demonstration, the feature sub-network trainer improved classification accuracy of a seven-layer convolutional neural network, trained with two thousand examples, from 75% to 83% by reusing the training examples.