SRAM Lookup Table for On-Device ML Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning techniques on portable, battery-operated devices face challenges due to high processing, data, and power resource consumption, often requiring offloading to remote networks, which can result in irrelevant data transmission, wasting bandwidth and raising privacy concerns.
Innovation Solution
A static random access memory (SRAM) circuit configured to store machine learning data in a lookup table, allowing consecutive access to selected SRAM cells, and a memory controller to select cells based on specific information, utilizing distributed arithmetic for efficient multiply and accumulate operations, reducing the need for complex multiplications and minimizing resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning operations are performed on portable devices, then processing capability is improved, but power consumption and resource usage increase
Solution Approach 1:
The patent segments the machine learning processing task into two distinct phases: a training phase that occurs offline on powerful systems, and an operating phase that runs on portable devices. This segmentation allows complex model training to be performed once, while the device only executes the lighter-weight inference operations, thereby improving processing capability on portable devices while keeping power consumption manageable.
Solution Approach 2:
The patent applies preliminary action by pre-training machine learning models offline before deployment to portable devices. The training phase prepares the models in advance, so that when the device operates, it only needs to perform inference using pre-computed models and weights. This preliminary preparation reduces the computational burden and power consumption during actual device operation.
2Device complexity
If machine learning operations are offloaded to remote networks, then processing power requirements are reduced, but data transmission bandwidth is wasted and privacy concerns increase
Solution Approach 1:
The patent extracts the computationally intensive training operations from the portable device and relocates them to remote training systems. Only the essential model parameters and weights are transferred to the device, while the bulk of data processing and model training remain on the server. This extraction reduces the device's processing power requirements while minimizing unnecessary data transmission, as only processed results rather than raw data are transmitted.
3Measurement precision
If complex machine learning models are deployed on portable devices, then recognition accuracy is improved, but resource consumption increases
Solution Approach 1:
The patent performs model optimization and compression as preliminary actions during the offline training phase. Complex models are trained on powerful systems with abundant resources, then optimized for deployment on portable devices. This preliminary preparation enables the device to run accurate models with reduced resource consumption, as the heavy lifting of model development and optimization occurs beforehand on more capable systems.
Data Source
AI summary
In a memory device, a static random access memory (SRAM) circuit includes an array of SRAM cells arranged in rows and columns and configured to store data. The SRAM array is configured to: store a first set of information for a machine learning (ML) process in a lookup table in the SRAM array; and consecutively access, from the lookup table, information from a selected set of the SRAM cells along a row of the SRAM cells. A memory controller circuit is configured to select the set of the SRAM cells based on a second set of information for the ML process.