Design and Implementation of High-Order CMAC Chip with On-Chip Learning Capability
碩士 === 國立成功大學 === 資訊工程學系研究所 === 85 === Albus所提出之小腦控制模式(CMAC: Cerebellar Model Articulation Controller),由 於具快速學習及高概括(generalization)兩迷人特性,已廣泛地使用在控制領域。然而矩 形的感知強度函數(receptive field function)卻造成不連續的函數近似。藉由引入B條 樣曲線函數(B-Spline function)至小腦控制模式中當作...
Main Authors: | , |
---|---|
Other Authors: | |
Format: | Others |
Language: | en_US |
Published: |
1996
|
Online Access: | http://ndltd.ncl.edu.tw/handle/00549162470803359466 |
id |
ndltd-TW-085NCKU0392010 |
---|---|
record_format |
oai_dc |
spelling |
ndltd-TW-085NCKU03920102015-10-13T12:18:05Z http://ndltd.ncl.edu.tw/handle/00549162470803359466 Design and Implementation of High-Order CMAC Chip with On-Chip Learning Capability 具On-Chip學習能力之高階小腦模式晶片設計與製作 Kao, Jen-Hwa 高建華 碩士 國立成功大學 資訊工程學系研究所 85 Albus所提出之小腦控制模式(CMAC: Cerebellar Model Articulation Controller),由 於具快速學習及高概括(generalization)兩迷人特性,已廣泛地使用在控制領域。然而矩 形的感知強度函數(receptive field function)卻造成不連續的函數近似。藉由引入B條 樣曲線函數(B-Spline function)至小腦控制模式中當作感知強度函數, Lane等人發展出 高階小腦控制模式,在函數逼近上可成功地消除不連續的函數近似。在色彩修正上,介於掃 描器與印表機之間的色彩映射是一種高度非線性映射,我們引用高階小腦控制模式來控制 其間的色彩映射關係,並完成一用於色彩修正之高階小腦控制模式晶片,然而在學習部份因 使用軟體達成,使得系統的效能大受影響。本篇論文中,我們達成學習演算法之硬體實作以 增進原有系統之效能。此外對於高階小腦模式在函數逼近及色彩修正問題上,我們對一些 系統參數作了詳盡的模擬分析,實驗結果顯示兩組相似的結果可藉由兩組Q/d比值相當的Q 與d而得,其中Q為輸入空間的分割數, d為B條樣取線的膨脹係數。這樣的結果可作為建構 一經濟的高階小腦控制模式在參數選取上的選擇方向。 The CMAC (Cerebellar Model Articulation Controller) proposed by Albus hasbee n widely used in the field of control engineering due to its attractive proper ties of fast learning speed and powerful generalization capability.The retangu lar receptive field functions, however, result in discontinuousfunction appoxi mation. Lane developed the higher-order CMAC neural network model by using B-S pline functions as receptive field functions, and provedthat the staircase fun ction approximation can be eliminated successfully byhigher-order CMAC. The co lor mapping between scanner and printer in color reproduction is highly nonlin ear. In this thesis, we employ higher-order CMAC to perform the color mapping. Moreover, a digital higher-order CMACchip for color correction has been devel oped. However, the system performanceis restricted because the learning algori thm is realized in software.Therefore, we further complete the hardware implem entation of CMAC learningalgorithm to speedup the original CMAC-based system. Besides, we analyze someparameters in higher-order CMAC model, and simulation results reveal that twosimilar results can be achieved by two pairs of Q and d , if their Q/d valuesare alike, where Q is the quantization level in input spa ce and d is the dilation constant of B-Spline function. It can be adopted as a guideline inthe choice of parameter values for one to construct a cost-effect ive CMAC system. Yau-Hwang Kuo 郭耀煌 --- 1996 學位論文 ; thesis 123 en_US |
collection |
NDLTD |
language |
en_US |
format |
Others
|
sources |
NDLTD |
description |
碩士 === 國立成功大學 === 資訊工程學系研究所 === 85 === Albus所提出之小腦控制模式(CMAC: Cerebellar Model Articulation Controller),由
於具快速學習及高概括(generalization)兩迷人特性,已廣泛地使用在控制領域。然而矩
形的感知強度函數(receptive field function)卻造成不連續的函數近似。藉由引入B條
樣曲線函數(B-Spline function)至小腦控制模式中當作感知強度函數, Lane等人發展出
高階小腦控制模式,在函數逼近上可成功地消除不連續的函數近似。在色彩修正上,介於掃
描器與印表機之間的色彩映射是一種高度非線性映射,我們引用高階小腦控制模式來控制
其間的色彩映射關係,並完成一用於色彩修正之高階小腦控制模式晶片,然而在學習部份因
使用軟體達成,使得系統的效能大受影響。本篇論文中,我們達成學習演算法之硬體實作以
增進原有系統之效能。此外對於高階小腦模式在函數逼近及色彩修正問題上,我們對一些
系統參數作了詳盡的模擬分析,實驗結果顯示兩組相似的結果可藉由兩組Q/d比值相當的Q
與d而得,其中Q為輸入空間的分割數, d為B條樣取線的膨脹係數。這樣的結果可作為建構
一經濟的高階小腦控制模式在參數選取上的選擇方向。
The CMAC (Cerebellar Model Articulation Controller) proposed by Albus hasbee
n widely used in the field of control engineering due to its attractive proper
ties of fast learning speed and powerful generalization capability.The retangu
lar receptive field functions, however, result in discontinuousfunction appoxi
mation. Lane developed the higher-order CMAC neural network model by using B-S
pline functions as receptive field functions, and provedthat the staircase fun
ction approximation can be eliminated successfully byhigher-order CMAC. The co
lor mapping between scanner and printer in color reproduction is highly nonlin
ear. In this thesis, we employ higher-order CMAC to perform the color mapping.
Moreover, a digital higher-order CMACchip for color correction has been devel
oped. However, the system performanceis restricted because the learning algori
thm is realized in software.Therefore, we further complete the hardware implem
entation of CMAC learningalgorithm to speedup the original CMAC-based system.
Besides, we analyze someparameters in higher-order CMAC model, and simulation
results reveal that twosimilar results can be achieved by two pairs of Q and d
, if their Q/d valuesare alike, where Q is the quantization level in input spa
ce and d is the dilation constant of B-Spline function. It can be adopted as a
guideline inthe choice of parameter values for one to construct a cost-effect
ive CMAC system.
|
author2 |
Yau-Hwang Kuo |
author_facet |
Yau-Hwang Kuo Kao, Jen-Hwa 高建華 |
author |
Kao, Jen-Hwa 高建華 |
spellingShingle |
Kao, Jen-Hwa 高建華 Design and Implementation of High-Order CMAC Chip with On-Chip Learning Capability |
author_sort |
Kao, Jen-Hwa |
title |
Design and Implementation of High-Order CMAC Chip with On-Chip Learning Capability |
title_short |
Design and Implementation of High-Order CMAC Chip with On-Chip Learning Capability |
title_full |
Design and Implementation of High-Order CMAC Chip with On-Chip Learning Capability |
title_fullStr |
Design and Implementation of High-Order CMAC Chip with On-Chip Learning Capability |
title_full_unstemmed |
Design and Implementation of High-Order CMAC Chip with On-Chip Learning Capability |
title_sort |
design and implementation of high-order cmac chip with on-chip learning capability |
publishDate |
1996 |
url |
http://ndltd.ncl.edu.tw/handle/00549162470803359466 |
work_keys_str_mv |
AT kaojenhwa designandimplementationofhighordercmacchipwithonchiplearningcapability AT gāojiànhuá designandimplementationofhighordercmacchipwithonchiplearningcapability AT kaojenhwa jùonchipxuéxínénglìzhīgāojiēxiǎonǎomóshìjīngpiànshèjìyǔzhìzuò AT gāojiànhuá jùonchipxuéxínénglìzhīgāojiēxiǎonǎomóshìjīngpiànshèjìyǔzhìzuò |
_version_ |
1716857163382521856 |