LOW POWER REAL TIME ELECTRONIC NEURON VLSI DESIGN USING SUBTHRESHOLD TECHNIQUE Young Jun Lee, Jihyun Lee, Y.B. Kim, J. Ayers , A.Volkovskii , A.Selverston , H.Abarbanel* , M.Rabinovich Electrical and Computer Engineering Department, Marine Science Center ,Northeastern University, Boston,MA Dept. of Physics SIO*, Institute for Nonlinear Science , University of California, San Diego, La Jolla, CA yjlee@ece.neu.edu, jlee@ece.neu.edu, ybk@ece.neu.edu, lobster@neu.edu ABSTRACT We discuss a VLSI electronic neuron circuit that imple- ments Hindmarsh and Rose neuron model. Magnitude and time scaling techniques are employed for 2V power sup- ply operation. A subthreshold operation technique and a single MOS resistor are used to minimize area and power consumption. Output bursts of the electronic neuron can be modulated dynamically by varying the input voltage level. The circuit is designed using 0.25 m CMOS standard pro- cess, and the total power dissipation is 163.4 watt. 1. INTRODUCTION Developing analog locomotion controllers using actual bio- logical neuron circuitry has much advantages over conven- tional digitized processing using a microcontroller. As the demand for real time adaptability to environment increases, an analog controller becomes more attractive than a digi- tal approach. It is very difficult to process large amounts of sensory input from natural environment with digital circuits. A digital approach requires substantial hardware resources such as analog-to-digital / digital-to-analog converter, adder and multiplier. Although digital controllers show high com- puting power, their output results strongly depend on the resolution of data, which requires higher power and more area. Analog circuits can process with virtually infinite res- olution and much smaller area if low noise is guaranteed, this makes it possible to process large quantities of sen- sory inputs in real time with low power operation using sub- threshold techniques. An adaptive analog controller consists of neuron and synapse circuits that mimic the biological nervous systems. The neu- ron’s behavior is modeled using Hindmarsh and Rose (HR) differential equations[1]. One approach has been to simu- late the neural bursting using an oscillator[2]. However, this is not an exact implementation of actual biological neuron’s behavior because the slow adaptation current component ( ) is assumed to be constant and the circuit is so reduced that it does not produce realistic bursts. Electronic neurons have been built using 3-dimensional HR model by Pinto, [3]. This circuit was designed with 15V using discrete components, leading to large neuron hardware size and sig- nificant power consumption. This paper presents a precise and low power VLSI implementation of HR neuron circuit using small silicon area. Output bursts of electronic neu- rons can be dynamically modulated by varying input voltage levels. This neuron design is implemented using 0.25 m CMOS technology with 2V power supply voltage. Section 2 explains the neuron model and Section 3 describes the methodology to design neuron circuit with emphasis on low power and small area. Section 4 shows the simulation re- sults followed by conclusion in Section 5. 2. NEURON MODEL This paper presents a low power electronic neuron imple- mentation of HR neuron model. Compared to other neu- ron models such as Fitzhugh-Nagumo or Hodgkin-Huxley, the HR model has several advantages. The mathematical equations are simpler than Hodgkin-Huxley model, which is based on global behavior rather than eletrophysiological process. Therefore, it requires less hardware resources and less design complexity. Also this model provides an accu- rate output frequency-input current relationship, which is not described by Fitzhugh-Nagumo. The equations of HR neuron model are given by [1]. (1) (2) (3) where, : membrane potential of neuron, : recovery current of neuron, : adaptation current of neuron, : applied current, and : the leftmost equilibrium point of the neuron model without adaptation. Coefficients of the Equation (1),(2) and (3) are: = 1, = 3, = 3.024, = 1.01, = 5.0128, = 0.0021,