AI Data Chunking via Codec Change Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI data communication technologies face challenges in maintaining low-latency communication due to the bulkiness of neural network models, leading to high bandwidth and power consumption, with redundant data transmission between devices, especially in distributed computing environments like split computing and federated learning.
Innovation Solution
Implementing a first and second codec technique to detect changes in AI data frames and generate data chunks based on these changes, reducing data transmission to only the modified portions, and using a universal tensor data protocol to efficiently transfer AI data between devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If neural network models with numerous parameters are used to achieve high accuracy in AI applications, then model accuracy is improved, but data transmission size and bandwidth consumption increase
Solution Approach 1:
The patent segments the AI data transmission process into multiple layers, where intermediate data from each layer is transmitted separately. This allows for selective compression and optimization at each layer, reducing the overall transmission size while maintaining model accuracy through progressive refinement of data chunks.
Solution Approach 2:
The patent dynamically adjusts transmission parameters such as data chunk size, compression level, and precision based on network conditions and computational requirements. By changing parameters like the number of bits used to represent intermediate data or the frequency of transmissions, the system optimizes the balance between accuracy and transmission efficiency.
2Speed
If intermediate data is transmitted frequently for real-time AI applications, then response time is improved, but power consumption and bandwidth usage increase
Solution Approach 1:
The patent implements periodic transmission of intermediate data chunks at optimized intervals rather than continuous transmission. By determining optimal transmission periods based on application requirements and network conditions, the system maintains real-time responsiveness while significantly reducing power consumption and bandwidth usage compared to continuous data streaming.
Solution Approach 2:
The patent transmits only the necessary partial data chunks that contain meaningful changes or updates, rather than transmitting complete intermediate data sets. This partial action approach ensures real-time application performance by sending only critical information, thereby reducing overall transmission volume and power consumption.
3Reliability
If redundant intermediate data is transmitted between devices in distributed computing, then data completeness is improved, but communication efficiency deteriorates
Solution Approach 1:
The patent extracts and transmits only the essential and non-redundant portions of intermediate data that are necessary for maintaining data completeness. By identifying and removing duplicate or redundant information before transmission, the system ensures that receiving devices obtain all necessary data for accurate reconstruction while improving communication efficiency through reduced transmission volume.
Data Source
Figure 1~3
Figure 4a
Figure 4b
AI summary
The present disclosure relates to methods and devices for managing communication of Artificial Intelligence (AI) data between devices in a computing environment. Detecting one or more sets of changes in data values associated with current data frame based on comparison between current data frame and previously generated data frame. Further, performing generation of one or more data chunks of an updated data frame, comprising one or more data segments based on the one or more sets of changes, based on one of, a first codec technique and second codec technique. Transmitting one or more data chunks to one or more receiver devices based on the respective data chunk satisfying a specified chunk size. Transmitting the one or more data chunks generated for updated data frame to one or more receiver devices based on respective data chunk satisfying a specified chunk size.