A Four-quadrant Floating-gate Synapse
نویسندگان
چکیده
We present a new type of pFET synapse; by degenerating the source, the oxide currents provide stabilizing feedback to the oating gate and to the drain. We present experimental measurements from a oating-gate synapse that simultaniously computes four-quadrant products of its input and weight values, and computes a four-quadrant correlation, typical of hebbian and backpropagation learning rules, between the input and drain voltages. Our four-quadrant synapse is built from two source-degenerated pFET synapses; by adding weak exponential feedback to the source of a oating-gate pFET synapse, we obtain a oating-gate synapse with unique dynamical properties. This four-quadrant synapse can become the fundamental building block of many continuoustime neural-network learning algorithms. Floating-gate devices have come to be seen not only as memory elements, but as continuous-time circuit elements operating a variety of timescales. This paper presents a oating-gate synapse that computes four-quadrant products of its input (Xi) and its weight value (Wi;j), and computes a four-quadrant correlation of the input and error signal (ei). Figure 1 shows a diagram of a continuously adapting neural array; the inset shows schematically the synapse computations for the ith row and for the jth column. This error signal could either be a function of the output, as in Hebbian or other unsupervised algorithms, or due to an error in the desired output, as in gradient-descent algorithms. Four-quadrant multiplications are essential in most adaptivelter and neural-network applications. Biological synapses use two-quadrant multiplications, but excitatory (positive multiplication) and inhibitory (negative multiplication) synapses are used in a cooperative fashion to achieve this functionality when needed [3]. Our four-quadrant oating-gate synapses, as well as our previous synapses [2], are well suited to large synaptic arrays because of ve important properties. First, in the absence of learning, the weight is stored permanently, because the charge leakage from a oating gate is negligible. Second, the synapse's output current is the four-quadrant product of the input signal with the synaptic weight. Third, because our synapse only requires two oating-gate transistors per synapse, the synapse still requires minimal area. Fourth, the synapse can dissipate a minimal amount of power, because it can operate at subthreshold current levels. Fifth, this synapse implements a four-quadrant learning rule for Synapse W Input2
منابع مشابه
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS { II : ANALOG AND DIGITAL SIGNAL PROCESSING 1 A Four - Quadrant Floating - Gate
| Four-quadrant synapse stuu In our rst treatments of single-transistor synapses ((1], 2]; elaborated in 3]), we presented the electron-tunneling, hot-electron{injection, and multiplicative behavior of these devices. We used a feedback connguration to characterize the tunneling and hot-electron injection phenomena in these synapses 3]. We also derived an eeective learning rule, but we did not c...
متن کاملSymmetric ultralow-voltage floating-gate rail to rail four quadrant analog multiplier
The floating-gate transistor can be used to design analog circuits, such as current mirrors [1], current scaler[2] and current multiplier [3, 4] and divider [4]. The multiple input floating-gate transistor can be used to design ultra lowvoltage “pseudo differential” pairs [1], hence ULV rail-to-rail amplifiers and analog multipliers are feasible using the floating-gate technique. In this paper ...
متن کاملAnalog Signal Processing Circuits Using Floating Gate MOS Transistors
Low voltage non-linear computational circuits useful for analog VLSI signal processing applications based on floating gate MOS transistors (FGMOSFETs) are presented. The FGMOS transistors operate in the saturation region. The variable equivalent threshold voltage (VT) of the FGMOS transistor is exploited in such a way to transform it to a simple MOSFET of zero VT. A bias circuit using a convent...
متن کاملLow Voltage Floating Gate MOS Transistor Based Four-Quadrant Multiplier
This paper presents a four-quadrant multiplier based on square-law characteristic of floating gate MOSFET (FGMOS) in saturation region. The proposed circuit uses square-difference identity and the differential voltage squarer proposed by Gupta et al. to implement the multiplication function. The proposed multiplier employs eight FGMOS transistors and two resistors only. The FGMOS implementation...
متن کاملCompact EEPROM-based Weight Functions
We are focusing on the development of a highly compact neural net weight function based on the use of EEPROM devices. These devices have already proven useful for analog weight storage, but existing designs rely on the use of conventional voltage multiplication as the weight function, requiring additional transistors per synapse. A parasitic capacitance between the floating gate and the drain o...
متن کامل