Hamming Processing Unit for Neural Network Image Code Comparison
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning neural networks face challenges in processing large datasets due to memory-intensive content like images or videos, which require significant memory and data transfer resources, leading to inefficiencies in processing speed and accuracy, especially when relying on conventional systems that retrieve data from local caches or storage devices.
Innovation Solution
The implementation of a system that enables 'edge' processing of image codes by generating Hamming image codes on IoT devices or memory devices, allowing for direct comparison of image codes at the memory devices themselves, using techniques like supervised semantics-preserving deep hashing (SSDH) and Hamming distance calculations, thereby reducing memory requests and increasing processing efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If data is retrieved from local caches or storage devices in conventional systems, then data access is possible, but processing speed and accuracy deteriorate due to memory-intensive operations and data transfer requirements
Solution Approach 1:
The patent segments the data processing function by separating image code generation from image data storage. IoT devices generate compact Hamming image codes locally, while the memory system stores only these codes. This segmentation eliminates the need to transfer and process large image datasets, dramatically improving processing speed while reducing memory and energy consumption.
Solution Approach 2:
The patent uses Hamming image codes as simplified copies of original images. Instead of storing and processing full image data, the system creates compact binary embeddings that preserve essential matching information. These code copies enable fast comparison operations without requiring access to the original large image files, resolving the contradiction between processing speed and resource consumption.
2Quantity of substance
If large datasets are stored and processed in conventional systems, then comprehensive data analysis is possible, but memory requirements and data transfer bandwidth increase significantly
Solution Approach 1:
The patent fundamentally changes the data representation parameter from full image pixels to Hamming binary codes. This parameter transformation reduces the data size from megabytes per image to bytes per image code, enabling the system to handle vastly larger datasets while consuming minimal memory and bandwidth resources. The semantic-preserving hashing ensures that despite this drastic parameter change, matching accuracy is maintained.
3Loss of time
If image comparisons are performed in conventional systems, then matching can be achieved, but latency increases due to data retrieval from local caches or storage devices
Solution Approach 1:
The patent applies preliminary action by generating Hamming image codes at the IoT device before data transmission. This preprocessing step converts images to compact binary representations locally, so that when data is transferred to the memory system, only the codes need to be compared. This eliminates the need for real-time image processing during comparison operations, dramatically reducing latency and improving processing efficiency.
Data Source
AI summary
Examples of systems and method described herein provide for the processing of image codes (e.g., a binary embedding) at a memory system including a Hamming processing unit. Such images codes may generated by various endpoint computing devices, such as Internet of Things (IoT) computing devices, Such devices can generate a Hamming processing request, having an image code of the image, to compare that representation of the image to other images (e.g., in an image dataset) to identify a match or a set of neural network results. Advantageously, examples described herein may be used in neural networks to facilitate the processing of datasets, so as to increase the rate and amount of processing of such datasets. For example, comparisons of image codes can be performed “closer” to the memory devices, e.g., at a processing unit having memory devices.