Neural Network Self-Learning via Perturbation and Critic Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial neural network systems lack the ability to learn from their own successes and failures, limiting their adaptability and creativity in application domains.
Innovation Solution
A self-learning neural network system that employs a perturbed network to generate ideas, a critic network for evaluation, and reinforcement learning to improve patterns based on feedback, mimicking the brain's process of generating and reabsorbing ideas as memories through cumulative cycles of experimentation and learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If pre-trained artificial neural networks are used in a tandem arrangement with one network perturbed to generate novel patterns and another as a critic, then the system can generate novel and potentially useful patterns, but the system lacks the ability to learn from its own successes and failures
Solution Approach 1:
The patent implements feedback mechanisms where the critic network evaluates patterns generated by the perturbed network and provides reinforcement signals. This feedback loop enables the system to learn from its successes and failures by adjusting network parameters based on evaluation results, directly addressing the inability of pre-trained networks to adapt through experience.
Solution Approach 2:
The system performs self-learning by autonomously generating patterns, evaluating them, and using the evaluation results to modify its own network parameters. This self-service capability allows the system to progressively improve its pattern generation without external intervention, developing adaptability through cumulative learning cycles.
2Adaptability or versatility
If successive cycles of experimentation and learning are implemented, then progressive adaptability and creativity are achieved, but the computational time and resources increase
Solution Approach 1:
The patent applies partial perturbation to the network parameters rather than complete retraining in each cycle. By introducing controlled amounts of randomness and performing incremental learning adjustments, the system achieves progressive adaptability without requiring exhaustive computational resources for each experimentation cycle.
Solution Approach 2:
The system performs preliminary pattern generation using the perturbed network before full evaluation and learning cycles. This preliminary action allows the system to quickly generate candidate patterns that can be selectively refined, reducing the overall computational time required for successive learning cycles while maintaining progressive adaptability.
Data Source
AI summary
A discovery system employing a neural network, training within this system, that is stimulated to generate novel output patterns through various forms of perturbation applied to it, a critic neural network likewise capable of training in situ within this system, that learns to associate such novel patterns with their utility or value while triggering reinforcement learning of the more useful or valuable of these patterns within the former net. The device is capable of bootstrapping itself to progressively higher levels of adaptive or creative competence, starting from no learning whatsoever, through cumulative cycles of experimentation and learning. Optional feedback mechanisms between the latter and former self-learning artificial neural networks are used to accelerate the convergence of this system toward useful concepts or plans of action.