Integrated Memory Array for Bayesian Neural Inference at the Edge
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning platforms operating with deterministic variables are susceptible to high decision risk due to the inability to express uncertainty in inputs and are inefficient for safety/security-critical domains, and edge-analytics face challenges with high data movement and limited resources.
Innovation Solution
A memory module system incorporating SRAM arrays with embedded high-speed random number generators and peripheral digital-to-analog converters, enabling Bayesian Inference by storing and computing statistical densities within the memory, minimizing data movement and optimizing computing effort.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If deterministic inference methods are used in ML platforms, then computational efficiency is maintained, but prediction confidence and uncertainty expression are lost
Solution Approach 1:
The patent combines Bayesian inference capabilities with edge device architecture by integrating random number generators, digital-to-analog converters, and memory arrays into a unified system. This merging enables uncertainty quantification and prediction confidence expression while maintaining edge device deployment feasibility through co-location of computational and storage resources.
2Loss of time
If edge-analytics is implemented with cloud-based inference, then network latency is reduced, but data movement and resource constraints increase
Solution Approach 1:
The patent implements a nested architecture where the edge device contains an integrated memory array and processing units that perform inference locally. This nested structure embeds full inference capabilities within the edge device, eliminating the need for cloud communication and thereby reducing both latency and energy consumption associated with data movement.
Solution Approach 2:
The edge device performs inference autonomously using its own integrated resources including local memory arrays, random number generators, and processing units. This self-service capability eliminates dependence on external cloud infrastructure, enabling real-time inference with minimal energy expenditure on data transmission.
3Measurement precision
If high precision computing is used in Bayesian inference, then prediction accuracy improves, but computing effort and resource requirements increase
Solution Approach 1:
The patent replaces traditional digital computing mechanisms with analog computing approaches using digital-to-analog converters and current-mode processing in the memory array. This substitution enables parallel processing of Bayesian inference computations, maintaining high prediction accuracy through analog precision while achieving improved computing throughput via parallelism.
Solution Approach 2:
The patent transitions from sequential digital computation to parallel analog processing by mapping inference computations onto the two-dimensional memory array structure. This dimensional transformation enables simultaneous execution of multiple computational operations, thereby improving throughput while maintaining the precision required for accurate Bayesian inference.
Data Source
AI summary
A memory module system for a high-dimensional weight space neural network configured to process machine learning data streams using Bayesian Inference and/or Classical Inference is set forth. The memory module can include embedded high speed random number generators (RNGs). The memory module is configured to compute, store and sample neural network weights by adapting operating precision to optimize the computing effort based on available weight space and application specifications.