Limits...
Criticality as a Set-Point for Adaptive Behavior in Neuromorphic Hardware.

Srinivasa N, Stepp ND, Cruz-Albrecht J - Front Neurosci (2015)

Bottom Line: Experiment, theory and recent models have shown that neuronal networks at criticality demonstrate optimal information transfer, learning and information processing capabilities that affect behavior.In this perspective article, we argue that understanding how large scale neuromorphic electronics can be designed to enable emergent adaptive behavior will require an understanding of how networks emulated by such hardware can self-tune local parameters to maintain criticality as a set-point.We believe that such capability will enable the design of truly scalable intelligent systems using neuromorphic hardware that embrace complexity in network dynamics rather than avoiding it.

View Article: PubMed Central - PubMed

Affiliation: Information and System Sciences Lab, Center for Neural and Emergent Systems, HRL Laboratories LLC Malibu, CA, USA.

ABSTRACT
Neuromorphic hardware are designed by drawing inspiration from biology to overcome limitations of current computer architectures while forging the development of a new class of autonomous systems that can exhibit adaptive behaviors. Several designs in the recent past are capable of emulating large scale networks but avoid complexity in network dynamics by minimizing the number of dynamic variables that are supported and tunable in hardware. We believe that this is due to the lack of a clear understanding of how to design self-tuning complex systems. It has been widely demonstrated that criticality appears to be the default state of the brain and manifests in the form of spontaneous scale-invariant cascades of neural activity. Experiment, theory and recent models have shown that neuronal networks at criticality demonstrate optimal information transfer, learning and information processing capabilities that affect behavior. In this perspective article, we argue that understanding how large scale neuromorphic electronics can be designed to enable emergent adaptive behavior will require an understanding of how networks emulated by such hardware can self-tune local parameters to maintain criticality as a set-point. We believe that such capability will enable the design of truly scalable intelligent systems using neuromorphic hardware that embrace complexity in network dynamics rather than avoiding it.

No MeSH data available.


(A) A typical configuration for interacting with the neuromorphic hardware, for instance when conducting a parameter search. Test software runs on a general purpose computer, which communicates with an FPGA over a USB connection. The connection allows software to upload networks to the chip, set hardware parameters, and perform spike-based input and output. (B) Flow chart detailing the parameter search process and its relation to each system component.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4664726&req=5

Figure 2: (A) A typical configuration for interacting with the neuromorphic hardware, for instance when conducting a parameter search. Test software runs on a general purpose computer, which communicates with an FPGA over a USB connection. The connection allows software to upload networks to the chip, set hardware parameters, and perform spike-based input and output. (B) Flow chart detailing the parameter search process and its relation to each system component.

Mentions: These deficiencies may be addressed if the chip can be set up to operate with criticality as a set-point. To achieve this, a high-level parameter search similar to the one described in Stepp et al. (2015) can be run on the hardware itself. Figure 2 depicts a typical setup, where a neuromorphic chip is installed on a test board, along with a supporting FPGA. A general purpose computer communicates with the FPGA via a USB connection, which enables software to configure the chip as well as send and receive spikes. Once a network is uploaded, parameters can be quickly set and re-set by modifying on-chip registers. A typical search requires less than 100 iterations of 300 s each, amounting to a worst-case runtime of approximately 8 h. This search would result in a configuration that could be set once, without requiring access to low-level parameters such as synaptic weights. Self-tuning criticality, again as shown in Stepp et al. (2015), would then ensure that the network maintained a useful level of activity without input-specific tuning. If parts of the hardware break or begin to function differently, we expect an amount of fault tolerance. Without respect to self-tuning criticality, neural networks of this sort are already relatively tolerant (Srinivasa and Cho, 2012). Beyond this intrinsic fault tolerance, the self-tuning aspect described here extends this capability. At some point, however, the network dynamics will become too different and the search will have to be repeated. The nature of this breaking point is not well understood, and is a subject for further study.


Criticality as a Set-Point for Adaptive Behavior in Neuromorphic Hardware.

Srinivasa N, Stepp ND, Cruz-Albrecht J - Front Neurosci (2015)

(A) A typical configuration for interacting with the neuromorphic hardware, for instance when conducting a parameter search. Test software runs on a general purpose computer, which communicates with an FPGA over a USB connection. The connection allows software to upload networks to the chip, set hardware parameters, and perform spike-based input and output. (B) Flow chart detailing the parameter search process and its relation to each system component.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4664726&req=5

Figure 2: (A) A typical configuration for interacting with the neuromorphic hardware, for instance when conducting a parameter search. Test software runs on a general purpose computer, which communicates with an FPGA over a USB connection. The connection allows software to upload networks to the chip, set hardware parameters, and perform spike-based input and output. (B) Flow chart detailing the parameter search process and its relation to each system component.
Mentions: These deficiencies may be addressed if the chip can be set up to operate with criticality as a set-point. To achieve this, a high-level parameter search similar to the one described in Stepp et al. (2015) can be run on the hardware itself. Figure 2 depicts a typical setup, where a neuromorphic chip is installed on a test board, along with a supporting FPGA. A general purpose computer communicates with the FPGA via a USB connection, which enables software to configure the chip as well as send and receive spikes. Once a network is uploaded, parameters can be quickly set and re-set by modifying on-chip registers. A typical search requires less than 100 iterations of 300 s each, amounting to a worst-case runtime of approximately 8 h. This search would result in a configuration that could be set once, without requiring access to low-level parameters such as synaptic weights. Self-tuning criticality, again as shown in Stepp et al. (2015), would then ensure that the network maintained a useful level of activity without input-specific tuning. If parts of the hardware break or begin to function differently, we expect an amount of fault tolerance. Without respect to self-tuning criticality, neural networks of this sort are already relatively tolerant (Srinivasa and Cho, 2012). Beyond this intrinsic fault tolerance, the self-tuning aspect described here extends this capability. At some point, however, the network dynamics will become too different and the search will have to be repeated. The nature of this breaking point is not well understood, and is a subject for further study.

Bottom Line: Experiment, theory and recent models have shown that neuronal networks at criticality demonstrate optimal information transfer, learning and information processing capabilities that affect behavior.In this perspective article, we argue that understanding how large scale neuromorphic electronics can be designed to enable emergent adaptive behavior will require an understanding of how networks emulated by such hardware can self-tune local parameters to maintain criticality as a set-point.We believe that such capability will enable the design of truly scalable intelligent systems using neuromorphic hardware that embrace complexity in network dynamics rather than avoiding it.

View Article: PubMed Central - PubMed

Affiliation: Information and System Sciences Lab, Center for Neural and Emergent Systems, HRL Laboratories LLC Malibu, CA, USA.

ABSTRACT
Neuromorphic hardware are designed by drawing inspiration from biology to overcome limitations of current computer architectures while forging the development of a new class of autonomous systems that can exhibit adaptive behaviors. Several designs in the recent past are capable of emulating large scale networks but avoid complexity in network dynamics by minimizing the number of dynamic variables that are supported and tunable in hardware. We believe that this is due to the lack of a clear understanding of how to design self-tuning complex systems. It has been widely demonstrated that criticality appears to be the default state of the brain and manifests in the form of spontaneous scale-invariant cascades of neural activity. Experiment, theory and recent models have shown that neuronal networks at criticality demonstrate optimal information transfer, learning and information processing capabilities that affect behavior. In this perspective article, we argue that understanding how large scale neuromorphic electronics can be designed to enable emergent adaptive behavior will require an understanding of how networks emulated by such hardware can self-tune local parameters to maintain criticality as a set-point. We believe that such capability will enable the design of truly scalable intelligent systems using neuromorphic hardware that embrace complexity in network dynamics rather than avoiding it.

No MeSH data available.