高性能矽視網膜感測器及具學習能力之細胞非線性網路晶片系統設計研究(III)

(1)

行政院國家科學委員會專題研究計畫成果報告

高性能矽視網膜感測器及具學習能力之細胞非線性網路晶

片系統設計研究(3/3)

計畫類別：個別型計畫計畫編號： NSC93-2215-E-009-016- 執行期間： 93 年 08 月 01 日至 94 年 10 月 31 日執行單位：國立交通大學電子工程學系暨電子研究所計畫主持人：吳重雨報告類型：完整報告報告附件：出席國際會議研究心得報告及發表論文國際合作計畫研究心得報告處理方式：本計畫可公開查詢

中華民國 94 年 12 月 5 日

(2)

高性能矽視網膜感測器及具學習能力之細胞非線性網路晶片系統

設計研究（結案）

The Researches on High-Performance Silicon Retina

Multi-Sensor and Learnable CNN Chips (結案)

計畫編號：NSC 93-2215-E-009-016

執行期限：93 年 8 月 1 日至 94 年 10 月 31 日

主持人：吳重雨國立交通大學電子工程系所

E-mail : [email protected]

計劃參與人員: 林俐如、陳勝豪、楊文嘉、吳諭、陳煒明、謝文芩一、中文摘要在具學習能力及比例式記憶體之細胞元化類神經晶片方面，藉著事先算出鍵值的平均，可提前決定出比例式記憶體之值，減少衰減操作的時間。在大鄰近層細胞非線性網路方面，我們也建立了新的架構，並且也在 MATLAB 軟體上驗證，成功的實現許多功能，目前用於實現此架構之晶片，已在 Hspice 模擬下驗證成功，並於 CIC 進行審查。我們研究了最新發表的生物視網膜細胞反應及其模型，利用此模型，將部分視網膜的功能成功的以電路實現。關鍵詞：矽視網膜、影像處理、運動感測、細胞化類神經網路、學習式類神經網路、細胞非線性網、大鄰近層細胞非線性網路 Abstract

As to the learnable CNN chips, we proposed a new learning scheme in which the average of the weights is evaluated first to determine the value of the ratio memory to avoid the elapsed time.

As to the large-neighborhood cellular nonlinear networks, the new structure is proposed. Meanwhile, this structure is also verified with MATLAB and many functions are successfully realized. Nowadays, the chip applied to implement this architecture is successfully verified in Hspice simulation

and the fabrication of this chip is under examination by CIC.

As to the silicon retina, we studied the newest publications regarding the response of biological retina and its model. Based on these data, we successfully developed a new silicon retina chip mimicking the functions of cells in biological retina.

Keywords: silicon retina、

pre-image-processing chip、motion detecting sensor、RMCNN、νBJT、 cellular nonlinear network、

large-neighborhood cellular nonlinear network 二、計畫緣由與目的我們已提出將細胞非線性網路當成聯想記憶體以用於學習、辨認及聯想樣本的觀念。各個細胞元之相互關係經由學習之後便儲存於比例式記憶體之中，並經由漏電來增強比例式記憶體之中超過平均值之鍵值，但是不可避免的是衰減操作的時間，因此本計畫的目的之一在使其能夠不須衰減操作。類神經網路係近年重要研究課程。我們在過去的研究計劃中，以載子在雙載子電晶體基極區域的傳輸現象，成功研發神經元雙載子電晶體(neuron-BJT)，並運用 neuron-BJT 實現矽視網膜及及大鄰近層連接細胞非線性網路 (Large Neighborhood Cellular Nonlinear Network, LNCNN)，並且提出新的大鄰近層架構，可以使得有更佳

(3)

的解析度，以及更容易控制，而有更多的功能，為了有更佳的解析度，勢必加大其陣列，因此我們也在此提出了新型電路，在加大解析度之餘，可以更節省功率的消耗。在以往的研究中，我們的矽視網膜晶片，乃在模仿此視網膜的 Photoreceptor、 Horizontal、及 Bipolar 這三種細胞三種細胞的反應。但事實上，以往對視網膜的研究，因受限於生物細胞訊號的量測技術尚未成熟，因此對視網膜細胞反應的瞭解並不完整。最近，在視網膜細胞反應的研究上，已有重大突破，許多視網膜細胞的反應已陸續被量測出來。本計畫在矽視網膜的研究目的，在學習並瞭解最新視網膜細胞反應的發現，並利用電路的方法將其實現。三、結果與討論在大鄰近層非線性網路方面，根據過去提出的架構如圖一(a)(b)，由於此架構的使用，將細胞模板 A 與細胞模板 B 分開來設計，而目前所改進的電路為圖二(a)，由於使用類比記憶體如圖二(b)，所以所需要的電路區塊，會比過去的架構少而且也比較簡單，同時每個區塊中所使用的電路可以更加簡單，而都可以實現大鄰近層非線性網路如圖三。目前我們使用的電路架構，為 neuron 圖四與記憶體圖五的部分，利用兩個元件的特性，使得在未開始運算時，不會有直流電流而可以達到節省功率，圖六為類神經突觸的部分，這部分我們將不同輸入方向的電路，分成兩部分，直接將其直流電流所經過的路徑打斷，因此可以不需要任何元件而可以節省功率。因此在電路待命的時候，此單一細胞元不需要功率消耗，而大鄰近層非線性網路的最主要目標，就是在於降低功率消耗以及電路面積，因此新的架構可以有更佳的效能且簡單，更能夠實現一個大型陣列的大鄰近層非線性網路通用機器。在比例式記憶體方面，我們由實驗晶片測試中文的一二四的辨識結果發現一二四皆能正確學習但在辨識四有兩個像素未能成功如圖八所示。此乃因為其中一個 T2 電路輸入為 1.5V 導致邏輯閘運作錯誤如圖九所示。改進方法為增加一條路徑使在圖形輸入時訊號流經此路徑如圖十所示。在矽視網膜晶片方面，根據最新視網膜細胞反應的發現，視網膜的 Photoreceptor 及 Horizontal 兩種細胞的反應可以用圖十一所示的模型去描述，τ是時間常數，D 是擴散常數，G 是訊號增益。根據圖十一的模型，我們設計出新型的矽視網膜電路，如圖十二所示為單一像素的電路，用以模擬 Photoreceptor 及 Horizontal 兩種細胞的反應。我們進一步用此電路建立一個解析度 32x32 的矽視網膜晶片，圖十三為晶片的照相圖。電路主要分成四個部分，分別實現生物模型中的四個部分。其中

Horizontal 的部分由電阻網路(resistive grid) 實現，Photoreceptor 的低通自回授的部分則利用 miller effect 加大等效電容以實現高時間常數的低通濾波器。此晶片模擬 Photoreceptor 及 Horizontal 兩種細胞的反應。經由晶片量測結果驗證此晶片在空間上與時間上的反應，均能和生物視網膜細胞相符(圖十四為晶片量測結果與生物量測的比較)。四、計畫成果自評在大鄰近層細胞非線性網路方面，從大架構的考量，有效的減少整個網路所需要的面積，並且進入內部電路的實現，也以朝向低功率消耗的電路設計方向進行，目標仍為了實現大鄰近層細胞非線性網路通用機器，而且在架構與電路的雙向進行下，更能夠達到更大的網路陣列，目前已經進行將此晶片下線至 CIC 進入其審查階段，

(4)

其 layout 如圖七，其 Post Simulation 結果如預期，可以將功率壓至非常低的程度，並且可以實現大鄰近層類神經網路的功能。在比例式記憶體方面，免衰減操作的觀念得到驗證但由於圖形輸入邏輯閘運作錯誤造成兩個像素錯誤。在矽視網膜方面，我們實現了一個新型的矽視網膜晶片。此晶片根據新近被測得的生物視網膜細胞反應而設計。此晶片模擬 Photoreceptor 及 Horizontal 兩種細胞的反應。經由晶片量測結果驗證此晶片在空間上與時間上的反應，均能和生物視網膜細胞相符，因此是一成功的設計。本年度計畫到目前為止執行成果極為豐碩，總計發表國際期刊論文八篇，國際學術會議論文二十篇，畢業學生四人，並列舉如下： 國際期刊論文

[1] Kuan-Hsun Huang, Li-Ju Lin, and Chung-Yu Wu, “The Analysis and Design of a CMOS Angular Velocity- and Directional-Selective Rotation Sensor with Retinal Processing Circuit,”

IEEE Sensors Journal, vol. 4, no. 6, pp.

845-856, December 2004.

[2] C.Y. Wu and K.H. Huang, “A CMOS focal-plane motion sensor with BJT-based retinal smoothing network and modified correlation-based algorithm,” IEEE Sensors Journal, vol. 2, no. 6, pp. 549-558, December 2002. [3] C.Y. Wu and C.H. Cheng, “A learnable

cellular neural network structure with ratio memory for image processing,” IEEE Transaction on Circuits and Systems I: Fundamental Theory and Applications, vol. 49, no. 12, pp. 1713-1723, December 2002.

[4] Yu-Chuan Shih and Chung-Yu Wu, “A New CMOS Pixel Structure for

Low-Dark-Current and Large-Array-Size Still Imager

Applications,” IEEE Trans. Circuits and

Systems I: Regular Papers, vol. 51, no.

11, pp. 2204-2214, Nov. 2004.

[5] Jui-Lin Lai and Chung-Yu Wu,

“Architectural Design and Analysis of Learnable Self-Feedback Ratio-Memory Cellular Nonlinear Network (SRMCNN) for Nanoelectronic Systems,” IEEE

Transactions on Vary Large Scale Integration (VLSI) Systems, vol. 12, pp.

1182-1191, Nov. 2004.

[6] Chung-Yu Wu and Cheng-Ta Chiang, "A Low-photocurrent CMOS Retinal Focal-plane Sensor with Pseudo-BJT Smoothing Network and Adaptive Current Schmitt Trigger for Scanner Applications," IEEE Sensors Journal, vol. 4, no. 4, pp. 510-518, Aug. 2004. [7] Chung-Yu Wu and Yu-Chuan Shih,

“Design, Optimization, and Performance Analysis of New Photodiode Structures for CMOS Active-Pixel-Sensor (APS) Imager Applications,” IEEE Sensors Journal, vol. 4, pp. 135-144, Feb. 2004.

[8] Chung-Yu Wu and Wen-Cheng Yen, "A new compact neuron-bipolar junction transistor (νBJT), cellular neural network (CNN) structure with programmable large neighborhood symmetric templates for image processing," IEEE Transaction on

Circuits and Systems I: Fundamental Theory and Applications, vol. 48, No. 1,

pp. 12-27, Jan. 2001. 國際學術會議論文

[1] Chung-Yu Wu and Yu Wu, “The Design of CMOS Non-Self-Feedback Ratio Memory Cellular Nonlinear Network without Elapsed Operation for Pattern Learning and Recogniton,” in

Proc. IEEE Int. Workshop on Cellular Neural Networks and their Applications, pp.282-285, 2005.

[2] J. L. Lai, Chung-Yu Wu, "A Learnable Self-Feedback Ratio-Memory Cellular Nonlinear Network (SRMCNN) with B Templates for Associative Memory Applications," in The 2004 IEEE

International Conference on Electronics, Circuits and Systems,

(5)

[3] S. H. Chen, Chung-Yu Wu, "A Low Power Design on Diffusive Interconnection Large-Neighborhood Cellular Nonlinear Network for Giga-scale System Applications," in The

2004 IEEE International Conference on Electronics, Circuits and Systems,

ICECS 2004, Dec. 2004.

[4] P. K. Lin, AF Lee, Chung-Yu Wu, Y. Chang, J. H. Liu. “The long-term electrophysiological recordings of macaca cyclopis from the subretinal microphotodiode array,” Taiwan Academy of Ophthalmology, Oct 2004. [5] P. K. Lin, Chung-Yu Wu, A. F. Lee, J.

H. Liu. “The electrophysiological recordings of macaca cyclopis from the implanted subretinal microphotodiode array,” XIII Afro-Asian Congress of Ophthalmology, Abstract number: 515. June 2004.

[6] Chung-Yu Wu, Felice Cheng, C. T. Chiang, and P. K. Lin, "A Low-Power Implantable Pseudo-BJT-based Silicon Retina with Solar Cells for Artificial Retinal Prostheses," in The 2004 IEEE

International Symposium on Circuits and Systems, ISCAS 2004, May 2004.

[7] P. K. Lin, A. F. Lee, Chung-Yu Wu, Y. Chang, J. H. Liu. “The Electroretinographic responses of macaca cyclopis from the implanted retinal prosthesis,” Asia Pacific Academy of Ophthalmology, Dec 2003. [8] Kuan-Hsun Huang, Li-Ju Lin, and

Chung-Yu Wu, “A CMOS Focal-Plane Rotation Sensor with Retinal Processing Circuit,” Proceedings of the 2003

European Solid-State Circuits Conference, pp. 129-132, Estoril,

Portugal.

[9] Chung-Yu Wu and Cheng-Ta Chiang, “A Low-photocurrent CMOS Retinal Focal-plane Sensor with Pseudo- BJT Smoothing Network and Adaptive Current Schmitt Trigger for Scanner Applications,” Proc. of the Second

IEEE International Conference on

Sensors, vol. 2, pp. 1147-1152, Oct.

2003.

[10] Wen-Chia Yang, Li-Ju Lin, and Chung-Yu Wu, “The Design of a Bionic Sensory Chip Based on the CNN Model Derived from the Mammalian Retina,”

Proc. of International Joint Conference on Neural Networks, 2003, pp. 371-375.

[11] Chiu-Hung Cheng, Sheng-Hao Chen, Li-Ju Lin, Kuan-Hsun Huang, Chung-Yu Wu, “A New Structure of Large-Neighborhood Cellular Nonlinear Network (LN-CNN),” Proc. of

International Joint Conference on Neural Networks, 2003, pp. 1497-1501

[12] Chung-Yu Wu; C.-Y. Hsieh; S.-H. Chen; B. C.-Y. Hsieh; C.-R. Chen, “Non-saturated binary image learning and recognition using the ratio-memory cellular neural network (RMCNN),” in proceedings of IEEE International

Workshop on CNNA2002, July, 2002,

pp. 624-629.

[13] C.-H. Cheng; Chung-Yu Wu, “The design of ratio-memory cellular neural network (RMCNN) with self-feedback template weight for pattern learning and recognition,” in proceedings of IEEE

International Workshop on CNNA2002,

July, 2002, pp. 609-615.

[14] T.-C. Tsai; M. S.; L.-J. Lin; C.-H. Cheng; Chung-Yu Wu, “A new

two-layer quantum-dot large-neighborhood cellular nonlinear

network (QLN-CNN) using quantum-dot cellular automata,” in proceedings of the IEEE Conference on

NANO2002, Aug. 26-28, 2002, pp.

355-357.

[15] Chung-Yu Wu, H.-C. Huang, L.-J. Lin, and K.-H. Huang, "A new pseudo-bipolar-junction-transistor

(PBJT) and its application in the design of retinal smoothing network," in the proceeding of IEEE International

Symposium on Circuits and Systems ISCAS 2002, May 2002, pp. 125-128.

(6)

[16] Chung-Yu Wu, H.-C. Huang, L.-J. Lin, and K.-H. Huang, "A new pseudo-bipolar-junction-transistor

(PBJT) and its application in the design of retinal smoothing network," in the proceeding of IEEE International

[17] Chung-Yu Wu; J.-L. Lai, “Improvement of pattern learning and recognition capability in ratio-memory cellular neural networks with non-discrete-type Hebbian learning algorithm,” in the proceeding of IEEE International

[18] L. J. Lin, C. H. Cheng, K. H. Huang, and Chung-Yu Wu, ”The Quantum-Dot Large-neighborhood Cellular Nonlinear

Network (QLN-CNN) in Nanotechnology,” in Proc. IEEE Conf.

Nanotechnology, IEEE-NANO 2001,

Maui, Hawaii, Oct. 2001, pp. 331-334. [19] Chung-Yu Wu, L. J. Lin, and K. H.

Huang, “A New Light-Activated CMOS Retinal-Pulse Generation Circuit Without External Power Supply For Artificial Retinal Prostheses,” in Proc.

Int. Conf. Electronics, Circuits, and Systems, ICECS’01, Malta, Sep. 2001,

Vol. II, pp. 619 –622.

[20] Chung-Yu Wu, K. H. Huang, and L. J. Lin, “The Design Of Cmos Real-Time Motion-Direction Detection Chip With Bjt-Based Silicon-Retina Sensors And Correlation-Based Motion Detection Algorithm,” in Proc. Int. Conf.

Electronics, Circuits, and Systems, ICECS’01, Malta, Sep. 2001, Vol. I, pp.

129–132. 畢業學生 楊文嘉、吳諭、陳煒明、謝文芩。 PAR PAL To Neuron C(0,1) and C(0,2) To Neuron C(0,-1) abd C(0,-2) To Neuron C(1,0) and C(2,0) To Neuron C(-1,0) and C(-2,0) PAIR PAIL PA3 PA1 Neuron C(0,0) PA D PA U PA 2 PA ID PA IU PAUL Neuron C(1,-1) PADL Neuron C(1,1) PADR Neuron C(-1,1) PAUR Neuron C(-1,-1) Output state voltageof C(0,0) Output of C(1,-1) Output of C(1,1) Output of C(-1,1) Output of C(-1,1) (a) IN PBL PBR PBIL PBIR PB3 PB1 PB 2 PB IU PB ID PB U PB D PBDR Neuron C(-1,1) PB0 PBDL PBUL PBUR Neuron C(1,1) To Neuron C(1,0) and C(2,0) To Neuron C(-1,0) and C(-2,0) To Neuron C(0,1) and C(0,2) To Neuron C(0,-1) and C(0,-2) Neuron C(-1,-1) Neuron C(0,0) Neuron C(1,-1) Input of C(0,0) Input of C(-1,1) Input of_C(1,1) Input of C(1,-1) Input of C(-1,-1) (b) 圖一、(a)神經元單一細胞模板 A 架構 (b) 神經元單一細胞模板 B 架構 PR1 PL2 PU1 PD2 PR2 PU2 PL1 PD1 PLU PLD PRD PRU PU2 NEURON Cell(0,0) Cell(1,0) Cell(1,1) Cell(0,1) Cell(-1,1) Cell(-1,0)

Cell(-1,-1) Cell(0,-1) Cell(1,-1)

(7)

NEURONCell(0,0) input external signal PS output S3 S5 neuron Memory S1 S2 PR1 PL2 PU1 PD2 PR2 PU2 PL1 PD1 PLU PLD PRD PRU PU2 NEURON Cell(0,0) Cell(1,0) Cell(1,1) Cell(0,1) Cell(-1,1) Cell(-1,0)

Cell(-1,-1) Cell(0,-1) Cell(1,-1)

(b) 圖二、(a)神經元單一細胞模板 A 與細胞模板 B 結合之架構(b)Neuron 區塊內部架構 Neuron C(-1,1) Neuron C(0,1) Neuron C(1,1) Neuron C(-1,0) Neuron C(0,0) Neuron C(1,0) Neuron C(-1,-1) Neuron C(0,-1) Neuron C(1,-1) Diffusive connection Direct connection Neuron C(0,-2) Neuron C(0,2) Neuron C(2,0) Neuron C(-2,0) r = 1 r = 2 圖三、舊架構與新架構所實現的大鄰近層 input Vdd Vss V1 V2 M1 M2 M3 M4 M5 M6 M7 M8 M9 M10 M11 Sp Sn Iin Iout M12 圖四、Neuron 電路圖 V dd V ss V 1 V 2 M 1 M 2 M 3 M 4 Iin Ioutp M 5 M 6 Ioutn S 2 S 1 圖五、記憶體電路圖 Vdd Vss Vbiasn Vbiasp M1 M2 M3 M4 M5 M6 Ipin Ipout Inin Inout 圖六、突觸電路圖

(8)

圖七、晶片 layout 圖圖八、四的辨識結果圖九、1.5V 導致邏輯閘運作錯誤圖十、改善後的電路圖十一、視網膜模型圖十二、矽視網膜電路

(9)

(a)單一像素 (b)整個晶片圖十三、晶片照像圖 (a) (b) 圖十四、(a)晶片量測結果(b)視網膜晶片量測結果與生物模型與生物實際量測比較，橫軸為時間，縱軸為位置，輸入影像為一閃爍的方塊，紅色長條標示為照光的空間與時間範圍

高性能矽視網膜感測器及具學習能力之細胞非線性網路晶片系統設計研究(III)

行政院國家科學委員會專題研究計畫 成果報告