Parameter Archival Storage for Neural Network Image Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neural-network based machine-learning models face challenges in interpretability, require extensive time and computational resources for training and optimization, and have cumbersome storage and management needs due to their complex architecture.
Innovation Solution
A system that includes an electronic query engine for constructing a model abstraction layer, versioning machine-learning neural-network based image processing models, and a parameter archival storage system for read-optimized compression storage, allowing for reduced storage and efficient management by storing only weight filter bits.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional neural-network based machine-learning models are used for complex image processing functions, then the ability to learn and model non-linear relationships is improved, but storage requirements and management complexity increase significantly
Solution Approach 1:
The patent segments the neural network model into two distinct parts: (1) a hierarchical linked architecture that defines the structural framework, and (2) weight filter bits that contain the actual learned parameters. This segmentation allows the architecture to be stored once and reused, while only the weight filter bits need to be updated and stored for different model versions, significantly reducing storage complexity while maintaining the full capability to model non-linear relationships.
Solution Approach 2:
The patent extracts and separates the weight filter bits from the complete model architecture for independent storage and management. By taking out only the essential weight parameters and storing them separately from the hierarchical architecture, the system reduces the burden of storing and managing complete model copies, while still enabling full model reconstruction when needed.
2Reliability
If complete neural network models are stored for versioning and mutation tracking, then model accuracy and functionality are preserved, but storage space requirements increase
Solution Approach 1:
The patent creates a simplified copy mechanism where only the weight filter bits (the essential learned parameters) are copied and stored for different model versions, rather than copying complete model architectures. This allows multiple model versions to be maintained with accurate weight parameters while using minimal storage space, as the hierarchical architecture definition is shared across all versions.
Solution Approach 2:
The patent discards redundant information by not storing complete model copies, and instead recovers the full model functionality by combining the stored weight filter bits with the hierarchical architecture definition when needed. This approach maintains model accuracy for versioning and mutation tracking while dramatically reducing storage space requirements.
3Ease of operation
If hierarchical linked architecture is stored for each model version, then model reconstruction capability is maintained, but storage efficiency decreases
Solution Approach 1:
The patent makes the hierarchical linked architecture universal by defining it once as a reusable framework that serves all model versions. This single architecture definition is multi-functional, providing the structural basis for reconstructing any model version when combined with the corresponding weight filter bits, thereby maintaining full reconstruction capability while maximizing storage efficiency.
Data Source
AI summary
Embodiments of the invention are directed to systems, methods, and computer program products for a parameter archival storage system for image processing models. The system is configured for read-optimized compression storage of machine-learning neural-network based image processing models with reduced storage by separately storing weight filter bits. The system is configured to construct weigh parameter objects associated with the plurality of neural network layers of an image processing model, such that the image processing model can be reconstructed from the weigh parameter objects. The system may discard the hierarchical linked architecture of the second image processing model and store the second image processing model at the at least one hosted model versioning system repository by storing only the weigh parameter objects.