Document Type : Original/Review Paper
Author
Faculty of Engineering, Electrical and Computer Department
Abstract
Biological observations indicate that forgetting is an inseparable part of the human learning system. Therefore, forgetting in learning algorithms is not necessarily destructive and can also be constructive. In practical implementations, due to limitations in storage capacity and the number of neurons in the network, only a limited number of training patterns can be learned, and subsequent patterns may cause destructive interference with previously learned ones. Consequently, for long-term learning, algorithms must incorporate a type of forgetting mechanism to create learning (storage) space for new training patterns. Therefore, to achieve success in the field of machine learning, a forgetting mechanism similar to that of the human brain is required. Forgetting is modeled as the loss of information from memory, and the necessity of such a mechanism becomes more pronounced in online learning, since the network must continuously update its weights. In this paper, the Active Learning Method, which is one of the widely used approaches, is employed. This method models the system based on the distribution of ink drops corresponding to training data. In this approach, the amplitude of ink drops on the planes remains unchanged, and no forgetting occurs, which is inconsistent with biological observations. The proposed mechanism improves adaptation in time-varying modeling tasks while maintaining performance comparable to conventional ALM on stationary data.
Keywords
- Fuzzy Inference System (FIS)
- Ink Drop Spread (IDS)
- Active Learning Method (ALM)
- Biological Neural Networks
Main Subjects
)