Model-aware minimal viable data transfer for neural network training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The time-consuming process of transmitting large training datasets for machine learning models, especially when training datasets are stored separately from the model, hinders efficient retraining and deployment in applications like object detection and motion detection.
Innovation Solution
A method and system that transmit a reduced training dataset, known as minimal viable data, which is reconstructed to match the input data format, allowing for efficient training while testing model accuracy and adjusting data reduction parameters as needed to ensure acceptable performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the entire training dataset is transmitted for model training, then the model training accuracy is ensured, but the data transmission time and storage requirements increase significantly
Solution Approach 1:
The patent extracts only the essential and relevant features from the complete training dataset to create a reduced training dataset. This extraction process identifies and retains minimal viable data that preserves model training accuracy while eliminating redundant information, thereby significantly reducing data transmission time and storage requirements without compromising the reliability of model training
Solution Approach 2:
The patent creates a simplified copy of the training dataset that contains only the necessary information for model training. This copied reduced dataset maintains the essential characteristics needed for accurate model training while being significantly smaller in size, thus resolving the contradiction between ensuring training accuracy and reducing transmission time
2Quantity of substance
If the entire training dataset is stored and transmitted, then complete training data is available, but the storage space and transmission bandwidth are consumed
Solution Approach 1:
The system extracts only the essential features and minimal viable data from the complete training dataset, removing redundant and non-essential information. This extraction maintains the necessary data completeness for effective model training while significantly reducing the quantity of data that needs to be stored and transmitted, thereby reducing storage space consumption
Solution Approach 2:
The patent applies parameter changes by transforming the training data into a reduced representation that preserves essential characteristics. This transformation modifies the data parameters to create a compact format that maintains training effectiveness while reducing the overall data volume, thus resolving the contradiction between data completeness and storage space requirements
3Productivity
If data reduction is applied to the training dataset, then the transmission time and storage requirements are reduced, but the model training accuracy may deteriorate
Solution Approach 1:
The patent carefully extracts only the most relevant and essential features from the training dataset, ensuring that the reduced dataset retains sufficient information for accurate model training. This selective extraction process maintains training accuracy while achieving significant data reduction, thus resolving the contradiction between transmission efficiency and model accuracy
Solution Approach 2:
The system applies intelligent parameter changes by transforming the training data into a reduced format that preserves critical training information. This transformation optimizes the data representation to maintain model training accuracy while improving transmission efficiency, effectively resolving the contradiction between these two parameters
4Loss of time
If the reduced training dataset is used for model training, then the transmission time is reduced, but the data format may not match the model input requirements
Solution Approach 1:
The patent introduces a data reconstruction module as an intermediary between the reduced training dataset and the model training process. This intermediary component reconstructs the reduced data into the required input format for the machine learning model, ensuring format compatibility while maintaining the benefits of reduced data transmission time. The reconstruction process bridges the gap between the compact reduced format and the model's expected input structure
Data Source
AI summary
Methods and systems for training a neural network include transmitting a first request for training data. The request includes information about the training data and information about a neural network model. A reduced training dataset is received that includes minimal viable data, responsive to the first request. A reconstructed training dataset is generated from the reduced training dataset. The model is trained using the reconstructed dataset.