擴散類神經網路在對數領域之超大型積體電路實現

Log-domain Implementation of Diffusion Network in VLSI

林士傑(Shih-Chieh Lin)

指導教授 ： 陳新

國立清華大學/電機資訊學院/電子工程研究所/碩士(2010年)

https://doi.org/10.6843/NTHU.2010.00505

摘要

在自然的環境中進行生物訊號的辨識，需要的是一個能夠對於雜訊影響有容忍度的處理系統，擴散網路在學習將演算法加入了雜訊與時變的元素，能夠正確且即時的反應生物訊號的變化，建立出在連續時間下有連續值變化的生物訊號分佈狀態。擴散網路實現於積體電路上是本研究主要的目標，然而超大型積體電路會面臨到晶片工作電壓的限制，導致在擴散網路中以電壓表示的神經元狀態變數與參數受到範圍侷限，因此能夠學習到難度較高的訊號並不容易。擴散網路在超大型積體電路的實現上利用對數領域的觀念設計電路，可以容忍電源供應電壓的下降並降低功率消耗，使得擴散網路處理訊號的行為將不受到限制。
對數領域使用在擴散網路上的電路設計觀念，主要是在MOS操作在次臨界區所擁有電流與電壓的指數關係，當狀態變數定義為電流時，可以對應出其節點電壓，這樣的指數關係式將原方程式轉換為對數領域的表示法，而新的方程式中的狀態變數將變為此節點電壓，原狀態變數也就表示說經過對數壓縮到節點電壓上，而經過對數領域的轉換將不會改變擴散網路應有的行為，狀態變數在電路上將具有數十倍的變動範圍。在實現於積體電路的過程中，我們會先對擴散網路進行數學模擬，取得學習不同訊號的參數範圍與設定，接著將數學上的數值轉換為對數領域在電路上的各種規格，建立出兩者的對應關係表，利用對數領域所使用到的各種電路完成整個擴散網路設計，在電路上重建出各種不同的訊號。本論文將擴散網路以晶片系統做為硬體實現，主要探討對利用對數領域的觀念實現擴散網路的電路設計，目的在於提升擴散網路的參數在積體電路的工作範圍，進而可以在電路上對更多樣的訊號做辨識處理。

關鍵字

擴散網路 ； 對數領域

並列摘要

The recognition of bio-medical signals needs to tolerate with the impact of environmental noise. the diffusion network (DN), proposed by Movellan in 2002, involves a stochastic process, which is capable of reflecting the variance of bio-medical signals in real-time. The reconstruction of continuous time, continuous valued signal is hence feasible. This research aims at the VLSI implementation of DN. As the technology evolves, the operating voltages of integrated circuits are further reduced, so are the dynamic ranges of DN variables. The limited operating range makes the learning of signals difficult. The log-domain concept circumvents the problem as well as reduces the power consumption of circuits. The dynamic range of the state variables in DN can operate over several decades without saturation.
The log-domain circuit exploits the exponential I-V relationships of MOS operated in sub-threshold region. The original state variables of the neurons are defined as the drain currents. The compressed states are derived from the gate voltage of the transistor. Though the log-domain translation, the diffusion network can be presented in an alternative form which is manipulating with the compressed states. Moreover, the state variables can operate over several decades without saturating the circuits. Before implementing the diffusion network with integrated circuits, we perform the numerical simulations to ensure the dynamic ranges of parameters of different signals. These values are then converted to the circuit domain so that the log-domain diffusion network can be implemented and plenty of signals can be learned with circuits. The dissertation implements the diffusion network from the hardware’s perspective. The log-domain concepts are employed in most circuit designs to enlarge the available dynamic ranges. As a result, versatile signals can be recognized with the proposed hardware.

並列關鍵字

Diffusion Network ； Log Domain

參考文獻

for partially observable diffusion processes: Theory and applications to neural

networks," Neural Computation, vol. 14, no. 7, pp. 1507-1544, 2002.

[2] H. Chen and A. F. Murray, "Continuous restricted Boltzmann machine with an

implementable training algorithm," IEE Proceedings-Vision Image and Signal

implementation of complex dynamics reaction-diffusion neural networks,"