increase performance. These transformations manifested as separable unique voltage
signals and are a consequence of the input signal entering a higher representational
space that is then used in reconstructing the desired target waveform. Unlike current
computational models of explicitly programmable algorithms, RC relies on systems
operating in a regime where they are able to ‘learn’ through experience,
circumventing the need for intelligent programming.
Generally, RC utilizes a randomly connected network, dubbed the ‘reservoir,’
composed of interacting elements called neurons with a topology based on mammalian brain neural networks. Information propagates through network connections
via electrical or electrochemical signals which are preferentially directed towards
neurons with the greatest connective strength. Neurons may be activated by the
incoming signal based on a set of learning rules causing them to undergo internal
modifications and excitatory transformation of the incoming signal. This stimulated
response of the neuron is then communicated through the neuron’s outgoing connections, thereby allowing the signal to percolate throughout the network. Computation is achieved by recording outputs from a few neurons assigned during
initialization and these neurons are trained to achieve the desired functionality.
Training resembles the evolutionary phenomenon observed in biological and natural
systems adapting due to environmental changes. External stimulation of the reservoir allows the system to independently evolve into a number of diverse neurons due
to signal propagation and local interactions which activate and modify neuron
properties. Clusters of neurons are capable of displaying distinct activity and emergent behaviors due to local interactions interplaying with signal percolation. Training selects a number of clusters desirable for computation and reinforces these
neurons for specialization into relevant mathematical processors. After the system
is properly trained using sample tasks, the neurons are passivated such that the
system retains the knowledge and experience for future computation. RC is both a
simple and elegant construction that avoids the need for absolute control over
programmable elements while capable of producing powerful processors due to
evolutionary concepts. Performance is controlled by neuron connectivity and distribution of strong and weak connections. Operational utility is thus dependent on the
reservoir’s statistical characteristics and global parameters, focusing on emergent
qualities of the network instead of individual elements.
The non-linear dynamics exhibited by the ASN device is uniquely positioned to
be exploited using the paradigm of reservoir computing [49, 63]. Interactions
between atomic switches in the network produce non-linear transformations of
input signals not present in Ohmic resistors or single switches. The additional
non-linear transformations of the input signal increases the diversity of the output
signals and thereby increase output separability and potential for computational
capability. In the reservoir computing approach, the input signal undergoes a
projection into a higher dimensional representation space, which can be thought of
as an expansion of the output into a sum of mathematical elements [64, 65]. Having a
rich collection of elements provides a large repertoire of possible transformations of
the input via mathematical representation vectors. The separability and utility of
multiple transformations enable the construction of mathematical algorithms by
Atomic Switch Networks for Neuroarchitectonics: Past, Present, Future
233
signals and are a consequence of the input signal entering a higher representational
space that is then used in reconstructing the desired target waveform. Unlike current
computational models of explicitly programmable algorithms, RC relies on systems
operating in a regime where they are able to ‘learn’ through experience,
circumventing the need for intelligent programming.
Generally, RC utilizes a randomly connected network, dubbed the ‘reservoir,’
composed of interacting elements called neurons with a topology based on mammalian brain neural networks. Information propagates through network connections
via electrical or electrochemical signals which are preferentially directed towards
neurons with the greatest connective strength. Neurons may be activated by the
incoming signal based on a set of learning rules causing them to undergo internal
modifications and excitatory transformation of the incoming signal. This stimulated
response of the neuron is then communicated through the neuron’s outgoing connections, thereby allowing the signal to percolate throughout the network. Computation is achieved by recording outputs from a few neurons assigned during
initialization and these neurons are trained to achieve the desired functionality.
Training resembles the evolutionary phenomenon observed in biological and natural
systems adapting due to environmental changes. External stimulation of the reservoir allows the system to independently evolve into a number of diverse neurons due
to signal propagation and local interactions which activate and modify neuron
properties. Clusters of neurons are capable of displaying distinct activity and emergent behaviors due to local interactions interplaying with signal percolation. Training selects a number of clusters desirable for computation and reinforces these
neurons for specialization into relevant mathematical processors. After the system
is properly trained using sample tasks, the neurons are passivated such that the
system retains the knowledge and experience for future computation. RC is both a
simple and elegant construction that avoids the need for absolute control over
programmable elements while capable of producing powerful processors due to
evolutionary concepts. Performance is controlled by neuron connectivity and distribution of strong and weak connections. Operational utility is thus dependent on the
reservoir’s statistical characteristics and global parameters, focusing on emergent
qualities of the network instead of individual elements.
The non-linear dynamics exhibited by the ASN device is uniquely positioned to
be exploited using the paradigm of reservoir computing [49, 63]. Interactions
between atomic switches in the network produce non-linear transformations of
input signals not present in Ohmic resistors or single switches. The additional
non-linear transformations of the input signal increases the diversity of the output
signals and thereby increase output separability and potential for computational
capability. In the reservoir computing approach, the input signal undergoes a
projection into a higher dimensional representation space, which can be thought of
as an expansion of the output into a sum of mathematical elements [64, 65]. Having a
rich collection of elements provides a large repertoire of possible transformations of
the input via mathematical representation vectors. The separability and utility of
multiple transformations enable the construction of mathematical algorithms by
Atomic Switch Networks for Neuroarchitectonics: Past, Present, Future
233