Neural Network Augmentation for Sequence Inference Stability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning and artificial intelligence networks struggle when deployed on sequences of images, as they lack context, leading to instability in inferred outcomes, and there are no existing systems or techniques to enlarge the context of such sequences without retraining the network.
Innovation Solution
A system comprising a network augmentation component that combines output data from different layers of an artificial neural network trained on static image data, and a network application component that employs the artificial neural network to infer outcomes from image instances in a sequence of images based on combined output data, thereby providing temporal and spatial context without retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a machine learning network trained on static images is deployed on sequences of images, then the network can process dynamic data, but the lack of temporal and spatial context leads to instability in inferred outcomes
Solution Approach 1:
The patent segments the sequence processing task by creating multiple instances of the network, each handling different temporal or spatial positions in the sequence. These segmented network instances process local context independently, then their outputs are aggregated to form the final inference, allowing the system to handle sequence data while maintaining stability through distributed processing
Solution Approach 2:
The patent implements nesting by embedding multiple network instances within a hierarchical structure where smaller network units process local sequence segments, and their combined outputs feed into larger aggregation layers. This nested architecture enables the network to capture both local temporal/spatial patterns and global sequence context, improving reliability without sacrificing adaptability
2Reliability
If the network is retrained to incorporate temporal and spatial context, then the inference stability improves, but the complexity and computational cost increase
Solution Approach 1:
The patent creates multiple copies of the existing trained network instance rather than retraining a single complex network. Each copy is deployed to process different aspects of the sequence data, and their results are combined. This copying approach achieves improved stability through ensemble processing while avoiding the computational burden of retraining
Solution Approach 2:
The patent makes the existing network universal by deploying identical network instances to handle multiple functions: processing different temporal positions, different spatial regions, or different feature aspects of the sequence data. This multi-functional deployment of simple network copies achieves the reliability benefits of complex contextual networks without the retraining complexity
3Loss of information
If the network size is increased to capture temporal and spatial context, then the context coverage improves, but the network compactness and efficiency are reduced
Solution Approach 1:
The patent merges the outputs of multiple simple network instances rather than using a single large network. By combining the results of several compact network copies that each process different sequence aspects, the system achieves comprehensive context coverage equivalent to a large network while maintaining the efficiency and compactness of individual small network units
Data Source
AI summary
Systems, computer-implemented methods, and computer program products that facilitate temporalizing and/or spatializing a machine learning and/or artificial intelligence network are provided. In various embodiments, a processor can combine output data from different layers of an artificial neural network trained on static image data. In various embodiments, the processor can employ the artificial neural network to infer an outcome from an image instance in a sequence of images based on combined output data from the different layers of the artificial neural network.