is optimized for specific tasks, SpiNNaker has implemented a versatile, tunable
neuromorphic architecture in the same vein as FPGAs at the cost of power efficiency. Both architectures have been designed for scalability and can accommodate
cascades of devices allowing for easier implementations of larger systems.
Qualcomm Technologies Zeroth Machine Intelligence platform has also developed a deep learning software development kit (SDK) which has brought the power
of deep learning to mobile devices. This breakthrough has removed the need to
connect to a cloud server to utilize the benefits of deep learning and gives mobile
phones the innate capability to perform numerous complex tasks like facial recognition, object tracking and natural language processing [43].
3.5 ASNs for Computing
ASN devices exhibit various properties associated with atomic switches and other
memristive systems [44], the latter defined as a system with its internal resistance
based on electric flux [45]. Such properties include but are not limited to a requisite
forming step and distributed, frequency-dependent hysteretic switching among a
collection of dynamically interacting elements. In addition, the functional topology
of ASNs has been shown to produce a diversity of complex behaviors, ranging from
distributed memory function to emergent critical dynamics similar to those found in
both fMRI/EEG of biological brains and multi-electrode array (MEA) studies of
neuronal populations [15].
Observations of power-law scaling in various device dynamics and a ‘fading
memory’ property of learned states have implicated as an essential component for
applications of reservoir computing (RC) using critical states. Initial progress in the
use of ASNs as nonlinear reservoirs capable of task performance in the RC paradigm
has shown through simulation and experimental implementation of a benchmark
task known as waveform generation.
Based on extensive studies of the dynamical response of ASNs (Fig. 4), these
devices have been identified as an ideal platform for hardware-based reservoir
computation [46]. ASN devices have shown, through both experiment and simulation [47–49], to be a viable platform for hardware-based RC toward applications in
pattern recognition, prediction and logic. Based on their ability to integrate, segregate, store and respond to external stimulus, the utility of ASNs as nonlinear
reservoirs has been demonstrated through implementation of multiple benchmark
tasks including: (1) waveform generation [48] and (2) various logic operations
(AND, OR, XOR) (see Sect. 3.2). The speed, density, and [50] scalability of the
ASN serve to overcome major hurdles in the RC paradigm.
Atomic Switch Networks for Neuroarchitectonics: Past, Present, Future
213
neuromorphic architecture in the same vein as FPGAs at the cost of power efficiency. Both architectures have been designed for scalability and can accommodate
cascades of devices allowing for easier implementations of larger systems.
Qualcomm Technologies Zeroth Machine Intelligence platform has also developed a deep learning software development kit (SDK) which has brought the power
of deep learning to mobile devices. This breakthrough has removed the need to
connect to a cloud server to utilize the benefits of deep learning and gives mobile
phones the innate capability to perform numerous complex tasks like facial recognition, object tracking and natural language processing [43].
3.5 ASNs for Computing
ASN devices exhibit various properties associated with atomic switches and other
memristive systems [44], the latter defined as a system with its internal resistance
based on electric flux [45]. Such properties include but are not limited to a requisite
forming step and distributed, frequency-dependent hysteretic switching among a
collection of dynamically interacting elements. In addition, the functional topology
of ASNs has been shown to produce a diversity of complex behaviors, ranging from
distributed memory function to emergent critical dynamics similar to those found in
both fMRI/EEG of biological brains and multi-electrode array (MEA) studies of
neuronal populations [15].
Observations of power-law scaling in various device dynamics and a ‘fading
memory’ property of learned states have implicated as an essential component for
applications of reservoir computing (RC) using critical states. Initial progress in the
use of ASNs as nonlinear reservoirs capable of task performance in the RC paradigm
has shown through simulation and experimental implementation of a benchmark
task known as waveform generation.
Based on extensive studies of the dynamical response of ASNs (Fig. 4), these
devices have been identified as an ideal platform for hardware-based reservoir
computation [46]. ASN devices have shown, through both experiment and simulation [47–49], to be a viable platform for hardware-based RC toward applications in
pattern recognition, prediction and logic. Based on their ability to integrate, segregate, store and respond to external stimulus, the utility of ASNs as nonlinear
reservoirs has been demonstrated through implementation of multiple benchmark
tasks including: (1) waveform generation [48] and (2) various logic operations
(AND, OR, XOR) (see Sect. 3.2). The speed, density, and [50] scalability of the
ASN serve to overcome major hurdles in the RC paradigm.
Atomic Switch Networks for Neuroarchitectonics: Past, Present, Future
213