Pixel sensor integrated neuromorphic VLSI system for real-time applications☆
Introduction
The parallel neuromorphic architectures in nature can provide further clues in the development of computational system models and their custom design methods. The cellular nonlinear (or neural) network (CNN) paradigm is often employed to address such functional analogies, although system model definitions may not exactly coincide. There are previous examples of the related parallelism in the literature, with their implementations. A particular category deals with spatiotemporal state excitable models realized as autonomous CNNs [5], [2], [15]. The more frequently highlighted applications of such models include artificial locomotion [3], [4], [1] and path optimization such as in [14]. Another vast CNN application area regards image processing, also with implemented examples [16], [17], [10]. Related VLSI design aspects and technology issues were also addressed in Refs. [11], [13]. In the addressing of certain implementation issues, for instance, minimization of the silicon area, custom design of specific neuromorphic system models can also provide various advantages. Recently, a bio-inspired system model was studied as such a specific CMOS CNN design by the authors, with verified system level functionality [9]. Certain applications were also shown numerically and through implementation, where the system model performed certain image processing tasks, such as image smoothing and edge detection enhancement. A coupling arrangement of the cells in the array also results in selective response in particular spatial directions, which was utilized for horizontal and vertical line detection [9]. Furthermore, the related network model also corresponds to the activator dynamics within the Fitzhugh–Nagumo neuroelectric model [6]. Accordingly, the cell states are excitable, which results in autowave propagation [7]. Implementation examples of Fitzhugh–Nagumo model exist in the literature, with top-down design methods as in [12]. It was shown, however, that a key nonlinear functionality in this model can be compactly realized using a custom design approach [8].
In this paper, we address a PN photodiode sensor integration with this specific neuromorphic system design. It is shown that the analog input states in the related 2D cell array can be programmed directly with visual patterns. The result is a pixel sensor equipped analog processor array, without a considerable layout overhead. A test chip in standard CMOS technology is implemented to validate the functionality of the sensor integrated design. The implementation can offer real-time computation performance as a compact stand-alone array or a building block, especially regarding excitable dynamical models. The custom design approach of such models can remarkably enhance the computational performance in related applications.
Section snippets
System model
The circuit-system level model includes a symmetric resistive coupled 2D processor array of nonlinear cells, which corresponds to a particular autonomous CNN architecture as depicted in Fig. 1. The dynamical state equation obtained from the current relation for each network node ij isHere, node voltage corresponds to the cell state, is a common bias current for each cell and is the nonlinear current response of the cell ij.
Sensor integration
In the design, the system inputs are defined as the initial cell voltage states within 2D array. Each cell employs Program/Execute modes of operation, where Execute logic input controls a synchronous switching between them. The Program mode establishes the initial cell states, whereas the Execute mode imposes the internal cell node () initial states. The steady-state response of the system to these states is considered as the processed output. For this operation, a previous design utilized
Static power
An important issue in large analog array implementations is the static power consumption. Herein, we focus on related consumption in the nonlinear array core at steady states . As the peripheral designs such as bias circuits can vary significantly, they are not included in this analysis. For the cubic cell circuit, the power consumption changes for the low and high stable states. In the low state , power dissipated per cubic cell circuit can be expressed as , neglecting BiasNL
Measured and numerical results
The functionality of a related test chip is verified with measurements, according to described operation cycle. We also present the measured I–V response of the seven transistor cubic cell circuit, which remained inherent in a previously implemented prototype. The resulting I–V curve is in Fig. 5. This response strongly agrees with the previously simulated results.
The bias voltages employed for chip measurements are , , , , and
Conclusion
The custom VLSI design of a neuromorphic system model is integrated with photodiode pixel sensors. A test chip is fabricated in CMOS; the related design and measured results are presented with comparisons. The particular collective system model can be utilized as a specific stand-alone processor or as a building block in the realization of various known excitable and neuroelectric models due to its parallelism. Furthermore, a seamless sensor integration with the design is possible without
Koray Karahaliloglu received the B.S., M.S. and Ph.D. degrees in Electrical Engineering from Bogazici University, Istanbul, Turkey, in 1993, 1996 and 2002, respectively. Between July 2001 and June 2004, he worked as a Post-Doctorate Fellow and then as a Research Assistant Professor in the Department of Electrical Engineering, University of Nebraska, Lincoln. Currently, he is an Assistant Professor in the Department of Electrical and Computer Engineering at Virginia Commonwealth University. His
References (17)
Impulses and physiological states in theoretical models of nerve membrane
Biophys. J.
(1961)Reaction–diffusion navigation robot control from chemical to VLSI analogic processors
IEEE Trans. Circuits Syst.—I: Regular Pap.
(2004)- et al.
Reaction–diffusion CNN algorithms to generate and control artificial locomotion
IEEE Trans. Circuits Syst.—I: Fundam. Theory Appl.
(1999) - et al.
CPG-MTA implementation for locomotion control
CNN wave based computation for robot navigation on ACE16K
- et al.
Autonomous cellular neural networks: a unified paradigm for pattern formation and active wave propagation
IEEE Trans. Circuits Syst.—I: Fundam. Theory Appl.
(1995) - et al.
Bio-inspired compact cell circuit for reaction–diffusion systems
IEEE Trans. Circuits Syst.—II: Express Briefs
(2005) - et al.
Compact MOS implementation of a reaction–diffusion CNN
Analog Integrated Circuits and Signal Processing
(2005)
Cited by (2)
ASIC implementation of a nonlinear dynamical model for hippocampal prosthesis
2018, Neural ComputationSelf-assembled ordered arrays of nanoscale germanium Esaki tunnel diodes
2011, Applied Physics Letters
Koray Karahaliloglu received the B.S., M.S. and Ph.D. degrees in Electrical Engineering from Bogazici University, Istanbul, Turkey, in 1993, 1996 and 2002, respectively. Between July 2001 and June 2004, he worked as a Post-Doctorate Fellow and then as a Research Assistant Professor in the Department of Electrical Engineering, University of Nebraska, Lincoln. Currently, he is an Assistant Professor in the Department of Electrical and Computer Engineering at Virginia Commonwealth University. His recent research interests include VLSI design of bio-inspired system models, collective system simulation, nanodevice computing applications and analog circuit-system design.
Patrick Gans received a B.S. degree in Electrical Engineering from the University of Nebraska-Lincoln in 2004. He is currently pursuing his M.S. degree in Electrical Engineering from the University of Nebraska-Lincoln. His research interests include cellular neural networks and real-time image processing. He is currently employed at STMicroelectronics as a Test Engineer for the micro, power, analog (MPA) product group in Carrollton, TX.
Nathan Schemm received the B.S. degree in Electrical Engineering from the University of Nebraska-Lincoln where he is currently working toward a Ph.D. Besides his work with cellular neural networks, his research interests include the VLSI integration of a wireless sensor network. He has worked on a new class of boron carbide diodes for neutron detection applications and is especially interested in the integration of such diodes into a wireless sensor network. His other research interests include CMOS imagers with focal plane compression and ultra-wideband technology.
Sina Balkir received the B.S. degree in Electrical Engineering from Bogazici University, Istanbul, Turkey, in 1987. He received the M.S. and Ph.D. degrees in Electrical Engineering from Northwestern University, Evanston, IL, in 1989 and 1992, respectively. Between 1992 and 1998, he was with the Department of Electrical and Electronics Engineering, Bogazici University, working as an Assistant and Associate Professor. Currently, he is with the Department of Electrical Engineering, University of Nebraska-Lincoln, where he serves as an Associate Professor. His research interests include CAD of VLSI systems, analog VLSI design automation and focal plane computation arrays for image processing.
- ☆
This work was supported by the US Office of Naval Research (ONR) under Grant N00014-01-1-0742.