Domain-Based Dendral Network Autonomous Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional artificial Neural Networks rely on back-propagation algorithms, making their learning capabilities dependent on human intelligence and limiting autonomous learning without human supervision.
Innovation Solution
The Domain Based Dendral Network (DBDN) employs Domain Based Tuning (DBT), where neurons naturally form groups (Domains) to accumulate knowledge, allowing continuous learning without algorithmic adjustments or human supervision, using input and tuning signals to adjust weights autonomously.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If back-propagation algorithm is used for training Neural Network, then learning capability is improved, but dependency on human intelligence and algorithmic complexity increases
Solution Approach 1:
The patent implements self-organizing maps that automatically learn and adapt without human supervision. The neural network performs unsupervised learning by naturally forming groups of neurons (domains) that accumulate knowledge about specific patterns, eliminating the need for back-propagation algorithms and human-guided training procedures
Solution Approach 2:
The patent changes the fundamental parameters of neural network training by replacing iterative back-propagation weight adjustments with domain-based autonomous learning. The system uses competitive learning where neurons compete to respond to input patterns, and winning neurons adjust their weights automatically based on pattern recognition needs rather than algorithmic directives
2Reliability
If back-propagation algorithm is used for training Neural Network, then desired results are obtained, but human supervision is required during training
Solution Approach 1:
The system implements autonomous learning through self-organizing maps where neurons automatically form functional domains and adjust their weights without human intervention. The network learns continuously by processing input patterns and naturally organizing its structure based on the statistical properties of the data, eliminating the need for human-supervised training sessions
Solution Approach 2:
The patent introduces dynamic domain formation where groups of neurons spontaneously organize themselves based on the input patterns they encounter. This dynamic reorganization allows the network to adapt its structure continuously without human supervision, with domains forming and evolving as new patterns are learned
3Ease of manufacture
If conventional Neural Network model is used, then implementation is straightforward, but autonomous learning without human supervision is nearly impossible
Solution Approach 1:
The patent segments the neural network into distinct functional domains, where each domain is responsible for learning specific patterns or classes of patterns. This segmentation into autonomous domains allows each group of neurons to learn independently without requiring global back-propagation, enabling distributed autonomous learning across the network
Solution Approach 2:
The patent introduces domain-based tuning as an intermediary mechanism between input patterns and weight adjustments. Instead of direct back-propagation, the system uses domain membership as a mediator to guide autonomous learning, where the domain structure itself facilitates the learning process without requiring external algorithmic control
Data Source
AI summary
A domain-based dendral network allows an artificial neural network to learn autonomously without any back-propagation algorithms or human supervision. An input bus transmits an input pattern to be analyzed into a neural network, which is composed of one of more neuron layers composed of multiple neurons, which perform analyze data propagated through the neural network with the aid of dendrons performing low-level signal analysis. An output bus collects the resulting output pattern from the neural network and sends the output pattern to a pattern comparator. The pattern comparator produces a tuning pattern by comparing the output pattern to a control pattern, and the timing pattern is sent to a tuning bus, which distributes the tuning pattern across the neural network according to a domain routing method. The use of the uniform routing method and the domain routing method facilitates the advantages of the present invention.