Biometric Classification System Using Shallow Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks require significant computational resources and large datasets for accurate classification, making them unsuitable for mobile applications with limited resources, such as IoT devices, and data-limited scenarios like surveillance or security access.
Innovation Solution
A scalable classification system architecture using integrated circuits with neural network computation units, employing Generative Adversarial Networks (GANs) for image feature enhancement and parallel processing with shallower neural networks to reduce computational and storage requirements, allowing for efficient classification of large groups of objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If very deep neural networks are used for classification, then classification accuracy is improved, but computational resource requirements increase
Solution Approach 1:
The patent divides the classification task into multiple stages: first a shallow neural network performs preliminary classification to identify potential matches, then deeper analysis is applied only to those cases. This segmentation allows the system to achieve high accuracy for relevant classifications while avoiding the computational cost of applying deep networks to all inputs.
Solution Approach 2:
The system applies computational resources partially - using shallow networks for most classifications and reserving deep network analysis only for cases where it is truly needed. This partial action approach maintains high accuracy for critical classifications while significantly reducing overall computational resource requirements compared to applying deep networks universally.
2Productivity
If cloud-based services are used to provide computation power, then classification performance is improved, but connectivity requirements and security risks increase
Solution Approach 1:
The patent implements a dynamic architecture that can adapt its operational mode based on connectivity conditions. The system is designed to function autonomously on mobile devices when disconnected, performing local classification with shallow networks, and can optionally enhance results with cloud-based deep network analysis when connectivity is available. This dynamic adaptability ensures continuous operation regardless of internet connectivity status.
Solution Approach 2:
The mobile device is equipped with local shallow neural network capabilities that enable it to perform classification independently without requiring cloud services. The device serves itself by maintaining essential classification functionality offline, reducing dependency on external infrastructure while preserving the option to leverage cloud resources when beneficial.
3Measurement precision
If large datasets are used for training, then classification accuracy is improved, but data storage requirements increase
Solution Approach 1:
The patent segments the training data into specialized subsets: shallow networks are trained on larger, more diverse datasets to handle general classification, while deep networks are trained on smaller, more specialized datasets for specific complex cases. This segmentation allows the system to achieve high overall accuracy without requiring all components to store and process extremely large datasets.
Solution Approach 2:
The system uses synthetic data generation and data augmentation techniques to create additional training examples from limited source data. By generating synthetic images and variations, the system effectively expands its training dataset without proportionally increasing storage requirements, enabling shallow networks to learn robust features from augmented data sets.
Data Source
AI summary
A biometric classification system includes a biometric capture system that captures a biometric identifier. A classifier includes a plurality of group classifiers. Each group classifier in the plurality of classifiers includes a group discriminator that determines, based on the captured biometric identifier, whether the biometric identifier belongs to a group of persons associated with the group discriminator, and includes a plurality of object discriminators. Each object discriminator is associated with a single person within the group of persons. The group discriminator determines whether the biometric identifier belongs to the group of persons. The object discriminator determines whether the biometric identifier belongs to the single person associated with the object discriminator.