Distributed Self-Organizing Map Grid for Scalable Neuromorphic Computing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing centralized implementations of Self-Organizing Maps (SOMs) suffer from scalability issues due to the Von Neumann bottleneck and massive interconnections, limiting their performance in terms of latency and hardware resources, especially in large-scale applications.
Innovation Solution
A neuromorphic computing system with a distributed architecture using a grid of locally connected cells, where each cell performs iterative computations and updates weights based on Euclidean distances and Manhattan distances, allowing for scalable hardware implementation through a MESH topology and time-multiplexed input processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If centralized implementation of SOM is used, then distance calculation can be concurrently executed for all neurons, but the system suffers from Von Neumann bottleneck and is not scalable in terms of latency
Solution Approach 1:
The patent segments the SOM architecture into distributed neuronal processing elements (NPEs) organized in a grid topology. Each NPE independently calculates distances to input vectors and identifies winning neurons locally, eliminating the centralized Von Neumann bottleneck. This segmentation enables parallel processing across multiple NPEs while reducing latency through distributed computation.
Solution Approach 2:
The patent introduces a spatial dimension by organizing NPEs in a two-dimensional grid topology rather than a centralized structure. This dimensional transformation allows simultaneous distance calculations across multiple NPEs operating in parallel, improving productivity while reducing the time loss associated with centralized data gathering and processing.
2Loss of time
If distributed implementation is used, then scalability in latency is improved, but massive interconnections are required which are not scalable
Solution Approach 1:
The patent implements local quality by limiting each NPE's connections to only its immediate spatial neighbors in the grid topology. Each NPE performs distance calculations and winner identification based solely on local information from adjacent NPEs, eliminating the need for massive all-to-all interconnections while maintaining scalability. This local processing approach reduces device complexity significantly.
3Productivity
If Euclidean distance calculation is performed iteratively across the grid, then scalability and reduced latency are achieved, but the computational process requires multiple clock edges
Solution Approach 1:
The patent employs periodic action by using a clock system with active clock edges to drive the iterative Euclidean distance calculation across the NPE grid. Each active clock edge advances the computation to the next NPE in sequence, enabling systematic propagation of distance calculations throughout the network. This periodic timing mechanism coordinates the iterative process while maintaining scalability.
4Adaptability or versatility
If time-multiplexed input processing is used, then high-dimensional data processing is supported, but the input processing becomes sequential rather than parallel
Solution Approach 1:
The patent applies preliminary action by pre-organizing high-dimensional input data into a structured format suitable for time-multiplexed processing. Each NPE receives and processes input vector components in a predetermined sequence across multiple clock edges, enabling systematic handling of high-dimensional data. This preliminary organization allows the system to maintain adaptability for high-dimensional inputs while managing the sequential processing requirement efficiently.
Data Source
AI summary
A neuromorphic computing system configured to be trained using unsupervised learning through distributed computing circuits. The neuromorphic computing system comprises an artificial neural network implemented as a grid of locally connected cells wherein each cell comprises hardware components for neural computing and storage, and is connected to its direct closest neighbors. The neuromorphic computing system comprises a clock system providing periodic active clock edges allowing in each cell to simultaneously and synchronously compute the neuron's Euclidean distance to the input, then compute the Best Matching Unit and the Manhattan distance to it in multiple clock cycles based on a time to Manhattan distance transformation, and finally update the neuron's weights.