Developmental Networks for Post-Selection-Free Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI methods suffer from Post-Selections, leading to weak generalization power and inefficiencies due to the need for human intervention in selecting systems based on luck rather than intrinsic mechanisms, resulting in high computational and manpower costs.
Innovation Solution
The Developmental Networks (DNs) approach, which avoids Post-Selections by using a brain-inspired framework that records and optimally evaluates performance across the lifetime of a neural network, allowing for incremental learning and self-generative vector representations, eliminating the need for human intervention in mapping functions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple systems are trained and post-selected based on performance, then the system selection accuracy is improved, but the computational cost and time consumption increase significantly
Solution Approach 1:
The patent extracts the performance evaluation from the training process itself by introducing a separate test set that is independent of the training data. This allows the training systems to be evaluated on their generalization capability rather than memorization, enabling selection based on true performance rather than training artifacts.
Solution Approach 2:
The patent introduces a test set as an intermediary between the training process and the final system selection. This test set acts as a mediator that objectively evaluates the generalization performance of multiple trained systems, allowing for fair comparison and selection without requiring extensive retraining or human intervention.
2Reliability
If multiple systems are trained and post-selected, then the reliability of system selection is improved, but the computational resources required increase
Solution Approach 1:
The patent applies partial action by training multiple systems with the same architecture and using a limited test set for evaluation. Rather than exhaustively searching all possible system configurations, the method selectively evaluates a representative subset of trained systems on the test set, achieving reliable selection with moderate computational investment.
3Adaptability or versatility
If handcrafted mapping functions are used to adapt systems to new environments, then the adaptability is improved, but the human intervention and time required increase
Solution Approach 1:
The patent enables systems to select and adapt to new environments autonomously through the post-selection process. When deployed in a new environment, the system is evaluated on the environment-specific test set, and the best-performing system is automatically selected without requiring human programmers to handcraft mapping functions or intervene in the adaptation process.
Data Source
AI summary
An analogy of Post-Selections is: “Somebody claims that his scheme provided a lottery ticket number that has won $1M, but he conceals that the scheme has spent 2 millions of lottery tickets of $1 each. The reported ticket is only the luckiest. The luckiest ticket was Post-Selected after the actual lottery test. The luckiest lottery ticket will not have the same luck next time.” Many machine learning methods suffer from Post-Selections, from neural networks, to reservoir computation, to swam intelligence to evolutionary computation. The numbers $1M, 2M and $1 and the chance to win in the analogy differ across different machine learning problems, but the nature of the flaw in the reports is basically the same. This invention presents a method that does not need any Post-Selections since it trains only one network that is computed in a closed form that corresponds to the most-probable network from training experience.