Efficient learning through biological neural network-Hebbian learning
Backpropagation provides new inspiration for neural network training, however, its biological rationality is still questionable. Hebbian learning is a completely unsupervised and feedback-free learning technology, which is a strong contender for biologically feasible alternatives. However, so far, i...
Saved in:
| 主要作者: | |
|---|---|
| 其他作者: | |
| 格式: | Thesis-Master by Coursework |
| 語言: | English |
| 出版: |
Nanyang Technological University
2022
|
| 主題: | |
| 在線閱讀: | https://hdl.handle.net/10356/155123 |
| 標簽: |
添加標簽
沒有標簽, 成為第一個標記此記錄!
|
| 總結: | Backpropagation provides new inspiration for neural network training, however, its biological rationality is still questionable. Hebbian learning is a completely unsupervised and feedback-free learning technology, which is a strong contender for biologically feasible alternatives. However, so far, it has neither achieved the high-precision performance of backpropagation nor a simple training process. In this dissertation, we have designed three neural networks based on Hebbian learning, namely PHN, MOR and WLAH. The three Hebbian learning networks are based on the improved Hebbian method, which mainly includes changing the weight update equation, introducing activation thresholds, and increasing the sparsity of the hidden layer. These can effectively implement the Hebbian method through a simple training program. In addition, the improved Hebbian rule reduces the number of training cycles from 1500 to 50. At the same time, it changes training from a two-step process to a one-step process to improve training dynamics. Nevertheless, the three Hebbian learning networks still achieve test performance comparable to backpropagation and advanced algorithms on the MNIST dataset. |
|---|