Neural Network Incremental Learning via Probabilistic Training Set Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Incremental learning in neural networks is computationally expensive due to the phenomenon of catastrophic forgetting, where older trained examples are forgotten as new data is introduced, especially in dynamically changing environments like vehicles, requiring a more efficient learning algorithm.
Innovation Solution
A system for training neural networks using probabilistic training set selection, where the probability of selecting a training example is based on its recency of exposure, with a weighted distribution that inversely relates to how often the example has been seen, allowing for efficient training set creation and minimizing catastrophic forgetting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the neural network trains on all available examples at each step to prevent catastrophic forgetting, then the reliability of learning is improved, but the computational cost and training time increase significantly
Solution Approach 1:
The patent extracts only the necessary information from the full training set by using a probability distribution to select a subset of examples for training. Instead of processing all examples, the system selects examples based on their probability weights, thereby reducing computational cost while maintaining the ability to prevent catastrophic forgetting through the weighted selection mechanism.
Solution Approach 2:
The patent applies local quality by assigning different probability weights to different training examples based on their characteristics and importance. Not all examples are treated equally; instead, each example has a specific probability of being selected for training, allowing the system to focus computational resources on the most important examples while still maintaining overall learning reliability.
2Productivity
If the neural network trains only on new examples to improve computational efficiency, then the productivity is improved, but catastrophic forgetting occurs where older trained examples are forgotten
Solution Approach 1:
The patent applies preliminary action by pre-calculating and maintaining probability distributions for training example selection before actual training occurs. The probability distribution is updated in advance based on new examples, ensuring that when training happens, the system already has a ready-to-use selection mechanism that balances efficiency with knowledge retention without requiring retrospective adjustments.
Solution Approach 2:
The patent implements feedback by continuously updating the probability distribution based on the neural network's performance and the examples it has encountered. This feedback loop ensures that the selection probabilities adapt to maintain both computational efficiency and knowledge retention, adjusting the training strategy dynamically rather than using a fixed approach.
3Manufacturing precision
If the training set size is increased to improve learning accuracy, then the manufacturing precision of the model is improved, but the use of energy and computational resources increase
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the probability distribution parameters based on the training data characteristics and model performance. Instead of using a fixed training set size, the system changes the selection probabilities to optimize the balance between training accuracy and computational energy consumption, adapting the effective training set composition rather than simply increasing raw data volume.
Data Source
AI summary
Disclosed are systems and methods to incrementally train neural networks. Incrementally training the neural networks can include defining a probability distribution of labeled training examples from a training sample pool, generating a first training set based off the probability distribution, training the neural network with the first training set, adding at least one additional training sample to the training sample pool, generating a second training set, and training the neural network with the second training set. The incremental training can be recursive for additional training sets until a decision to end the recursion is made.