Binary Neural Network Beacon Segmentation for Message Recall

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Hopfield neural networks have limited learning diversity and recall capacity, restricting their applications due to a small upper bound on the number of independent patterns they can learn, which limits their interest and effectiveness in message recognition and discrimination.

Innovation Solution

A binary neural network architecture is introduced, where messages are segmented into sub-messages and processed using binary beacons and connections, allowing for increased learning diversity and recall power, particularly in the presence of erasure, and offering high discrimination between valid and invalid messages through parsimonious coding and maximum likelihood decoding.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If Hopfield neural networks are used for message learning and recall, then the network can perform associative memory functions, but the learning diversity and recall capacity are limited due to a small upper bound on the number of independent patterns

Engineering Contradiction:
Improvelearning diversityVSAvoidnumber of independent patterns
Core Design Contradiction:
Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The message is divided into B sub-messages, with each sub-message processed by a dedicated block of l beacons. This segmentation allows the network to learn M = l × B independent patterns, significantly increasing learning diversity compared to traditional Hopfield networks where all neurons compete for the same pattern space.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The invention transitions from a single-layer Hopfield network to a multi-block architecture where beacons are organized into B blocks, each handling a sub-message. This dimensional expansion from a single neuron pool to multiple organized blocks increases the capacity to store and recall independent patterns.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Quantity of substance

If more independent patterns are learned to increase learning diversity, then the recall capacity improves, but the network complexity and difficulty of decoding increase

Engineering Contradiction:
Improvenumber of learned messagesVSAvoidnetwork architecture complexity
Core Design Contradiction:
Quantity of substanceVSDevice complexity

Solution Approach 1:

Each block of beacons is specialized to process a specific sub-message, creating local quality in the network architecture. This organization simplifies decoding because each block independently handles its sub-message with l beacons, avoiding the global complexity that would arise from having all M patterns compete across a single neuron pool.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

By segmenting the network into B blocks, each with l beacons handling one sub-message, the invention decomposes the complex decoding problem into B simpler independent problems. The total capacity M = l × B is achieved without the exponential complexity growth that would occur in a monolithic network.

Inventive Principle:
Principle #1Segmentation

3Device complexity

If traditional Hopfield networks are used, then the network structure is simple, but the discrimination capability between valid and invalid messages is insufficient

Engineering Contradiction:
Improvenetwork structure simplicityVSAvoiddiscrimination capability
Core Design Contradiction:
Device complexityVSReliability

Solution Approach 1:

The segmentation into B blocks with l beacons each provides a structured framework that enhances discrimination capability. Each block's dedicated beacons create stronger, more reliable associations for their specific sub-messages, improving the network's ability to distinguish valid learned messages from invalid ones while maintaining manageable complexity.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP2609545B1Devices for learning and/or decoding messages using a neural network, learning and decoding methods, and corresponding computer programs
Publication Date: 2014.10.08 INST TELECOM TELECOM BRETAGNE
  • EP2609545B1 patent drawingFigure 1~3
  • EP2609545B1 patent drawingFigure 4~6
  • EP2609545B1 patent drawingFigure 7~10

AI summary

The invention relates to a learning and decoding technique for a neural network. The technique involves using a set of neurons, referred to as beacons, wherein said beacons are binary neurons capable of assuming only two states, i.e. an on state and an off state, said beacons being distributed in blocks, each of which includes a predetermined number of beacons, each block of beacons being allocated for the processing of a sub-message, each beacon being associated with a specific occurrence of said sub-message. The learning involves using: a means for splitting a message to be learned into B sub-messages to be learned, where B is greater than or equal to two; a means for activating, for a sub-message to be learned, a single beacon in each block to be in the on state, all of the other beacons of said block being in the off state; a means for creating connections between beacons, activating, for a message to be learned, connections between the on beacons of each of said blocks, said connections being binary connections capable of assuming only a connected state and a disconnected state.