Presentation is loading. Please wait.

Presentation is loading. Please wait.

Neural network (II) — HNN Hopfield Neural Network Date : 2002/09/24 Present by John Chen

Similar presentations


Presentation on theme: "Neural network (II) — HNN Hopfield Neural Network Date : 2002/09/24 Present by John Chen"— Presentation transcript:

1 Neural network (II) — HNN Hopfield Neural Network Date : 2002/09/24 Present by John Chen E-mail : phd9008@cs.nchu.edu.tw

2 2 Outline Preliminaries Introduction HNN algorithm Application & Researh Topic Conclusion

3 3 Preliminaries Neural Networks are built of neurons and their connections The characteristics of Neural Network Learning capability The capability of Storage Fault tolerance The capability of induce Pallel processing

4 4 Preliminaries(cont) Hebb learning rule Question : How to learn ? Where to keep memory? Hebb proposed learning rule in 1949 Learning is just a local appearance, it is correlated with the excited degrees between connected neurons It is also called correlated learning rule dW ij /dt = S j X j

5 5 Preliminaries(cont) Fig Computing Model of Neuron

6 6 Introduction Concept of Hopfield Neural Network Proposed by J. Hopfield in 1982 Provide the base of research theory Graph of Hopfield Neural Network

7 7 Introduction (cont) The properties of HNN Parallel input, Parallel output Operation process divide into two part  Memorizing process  Remembering process IN memorizing process  Update weights by Hebb learning rule  ∆W ij = ηX i X j IN remembring process  Output a result most similar to memorizing example by calculating

8 8 Introduction (cont) Two type of HNN Discrete : (1 or -1), (0 or 1) Continuous : real value between 0,1 The cost function of HNN X i : status value of i’th neuron X j : status value of j’th neuron W ij : connection weight between i’th & j’th neuron θ j : bias value of j’th neuron

9 9 HNN Algorithm Algorithm of memorizing Step 1 : set network parameters Step 2 : read connection weights set W ij = X i p X j p and W ii = 0

10 10 HNN Algorithm(cont) Algorithm of remembering Step 1 : set network parameters Step 2 : read connection weights Step 3 : Input initial vector X Step 4 : Calculate new vector X net i = W ij X j 1 if net i > 0 X i new = X i old if net i = 0 - 1 if net i < 0 Step 5 : repeat until network converge

11 11 Application & Researh Topic 時空型霍菲爾類神經網路於鼻咽部復發腫瘤之偵 測 — TAAI 2001 作者:張傳育 樹德科技大學 資訊工程系 提出立體時空型霍菲爾類神經網路 (SHNC) 來偵測鼻咽部復發腫瘤 SHNC 結合動態影像的時空資訊及像素點間的結構資訊, 對每個像素點作分類;可有效過濾影像中的雜訊 採用了競爭式的學習法則,加快了網路收斂的速度 霍菲爾類神經網路 (HNN) 是屬於非監督式學習網路,免除 了事先訓練網路的麻煩 經實驗可知 SHNC 所偵測的結果比 K-means, PCA 等方法 來得正確有效率

12 12 Application & Researh Topic(cont) Neural Networks for Visual Cryptography — with Examples for complex Acess Schemes Author : Suchen Chiang, Tai-Wen Yue Tatung University Proposed Q ’ tron NN Model Derived from HNN Design Energies function  For Halftoning  For Restoration  For (2,2) Visual Cryptography Can be extend for (2,2) Visual Cryptography but in another paper

13 13

14 14 Application & Researh Topic(cont) Reseach Reference Journal  Neural Network  IEEE Trans. on Neural Network  IEEE Trans. On system, Man, and Cybernetics  IEICE Trans. On Information and system Conference  ICNN 1987 ~ 1988 ( 神經網路國際研討會 )  IJCNN 1989 ~ ( 神經網路國際聯合研討會 ) : IEEE and INNS  TAAI 人工智慧與應用研討會 ( 台灣 )

15 15 Application & Researh Topic(cont) Research Direction Read and Reference Preprocess Model Modify

16 16 Conclusion Research like the learning rule of Neural Network If I have more time or more resource


Download ppt "Neural network (II) — HNN Hopfield Neural Network Date : 2002/09/24 Present by John Chen"

Similar presentations


Ads by Google