Distributed cellular computing system and method for neural-based self-organizing maps

Patent metadata

Field	Value
Publication number	US-12475360-B2
Application number	US-201917413342-A
Country	US
Kind code	B2
Filing date	Dec 11, 2019
Priority date	Dec 12, 2018
Publication date	Nov 18, 2025
Grant date	Nov 18, 2025

Abstract

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.

First claim

The invention claimed is:

1 . A neuromorphic computing system configured to be trained using unsupervised learning, the neuromorphic computing system comprising a Self-Organizing Map-like (SOM-like) artificial neural network implemented as a (N×M) grid of locally connected cells, each cell having initial weights and being connected to its direct closest neighbors, the neuromorphic computing system comprising a clock system providing periodic active clock edges, the neuromorphic computing system further comprising circuits configured to simultaneously and synchronously in each cell:
at a first active clock edge:
(a1) receiving a same multidimensional input data as the initial weights;
(b1) calculating a Euclidean distance value between the initial cell weights and said multidimensional input data; and
(c1) setting an initial Manhattan distance value;
iteratively at each of the next (N+M−2) active clock edges:
(d1) comparing the Euclidean distance value calculated at the previous active clock edge to the Euclidean distance values of the direct closest neighbors, to determine a minimum Euclidean distance value; and
(e1) if the minimum Euclidian distance value is different from the Euclidean distance value of said cell: updating the Euclidean distance value of said cell by the minimum Euclidean distance value, and updating a Manhattan distance value of said cell to a value represented by the number of clock cycles since the iterative process; and
at a next ((N+M−2)+1) active clock edge:
(f1) updating the cell weights using an unsupervised learning rule.

2 . The neuromorphic computing system of claim 1 , further comprising a data acquisition layer configured for acquiring raw data, and a pre-processing layer configured for extracting characteristic features from the acquired raw data and providing a multidimensional input data as a vector representation of the extracted characteristic features.

3 . The neuromorphic computing system of claim 1 , wherein the grid is organized in a mesh topology.

4 . The neuromorphic computing system of claim 1 , wherein the SOM-like artificial neural network is one of athe Kohonen SOM (KSOM), athe Dynamic SOM (DSOM), or athe Pruning Cellular SOM (PCSOM) artificial neural network.

5 . A Field Programmable Gate Array (FPGA) comprising the neuromorphic computing system of claim 1 .

6 . An Application Specific Integrated Circuit (ASIC) comprising the neuromorphic computing system of claim 1 .

7 . A method for operating a neuromorphic computing system configured to be trained using unsupervised learning, the neuromorphic computing system comprising a Self-Organizing Map-like (SOM-like) artificial neural network implemented as a (N×M) grid of locally connected cells, each cell having initial weights and being connected to its direct closest neighbors, the neuromorphic computing system comprising a clock system providing periodic active clock edges, the method comprising simultaneously and synchronously in each cell, the steps of:
at a first active clock edge:
(a1) receiving a same multidimensional input data as the initial weights;
(b1) calculating a Euclidean distance value between the initial cell weights and said multidimensional input data; and
(c1) setting an initial Manhattan distance value;
iteratively at each of the next (N+M−2) active clock edges:
(d1) comparing the Euclidean distance value calculated at the previous active clock edge to the Euclidean distance values of the direct closest neighbors, to determine a minimum Euclidean distance value; and
(e1) if the minimum Euclidian distance value is different from the Euclidean distance value of said cell: updating the Euclidean distance value of said cell by the minimum Euclidean distance value, and updating a Manhattan distance value of said cell to a value represented by the number of clock cycles since the iterative process; and
at a next ((N+M−2)+1) active clock edge:
(f1) updating the cell weights using an unsupervised learning rule.

8 . The method of claim 7 , further comprising, after step (d1), computing the Manhattan distance of each neuron to a Best Matching Unit, wherein the Best Matching Unit is a cell in the grid having the smallest Euclidian distance value, the Manhattan distance value being computed by applying a time to Manhattan distance transformation rule where a propagation time is equivalent to the Manhattan distance in a grid-shaped architecture.

9 . The method of claim 7 , further comprising a data acquisition layer configured for acquiring raw data, and a pre-processing layer configured for extracting characteristic features from the acquired raw data and providing a multidimensional input data as a vector representation of the extracted characteristic features.

10 . A non-transitory computer readable medium having encoded thereon a computer program comprising instructions for carrying out the steps of the method according to claim 7 when said computer program is executed on a computer.

11 . The method of claim 7 , wherein the grid is organized in a mesh topology.

12 . The method of claim 7 , wherein the SOM-like artificial neural network is one of a Kohonen SOM (KSOM), a Dynamic SOM (DSOM), or a Pruning Cellular SOM (PCSOM) artificial neural network.

Assignees

• Centre Nat Rech Scient
• Univ Cote Dazur Uca
• Ecole Nat Superieure De Lelectronique Et De Ses Applications Ensea
• Cy Cergy Paris Univ

Inventors

• Miramond Benoît
• Rodriguez Laurent
• Khacef Lyes

Classifications

• G06N3/088

  Non-supervised learning, e.g. competitive learning · CPC title

• G06N3/049

  Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs · CPC title

• G06N3/063 Primary

  using electronic means · CPC title