7 Current-Induced Dynamics of Chiral Magnetic Structures
175
size scale, and manipulability, magnetic textures could play an important role in the
development of such novel computational technologies [98, 99].
Magnetic artificial neural networks. The vast progress within the field of artificial
intelligence is mainly based on the widely enhanced available hardware power, while
most of the concepts have been suggested already a few years ago. So as with deep
artificial neural networks, which nowadays are widely used for different types of AI
applications. However, so far they are mostly performed on the existing hardware
which, due to the classical segmentation in computational units and storage, are
not optimally suited for these types of applications as their power consumption
shows. Instead, alternative architectures which adjust to the deep neural network
structure are proposed, with a focus of creating their central components, i.e., artificial
synapses and neurons, in hardware. There are also several suggestions for magnetic
neuromorphic computing [99, 100] and how to implement artificial neurons [101] and
synapses [102, 103]. In particular, memristors, [104] i.e., devices whose resistance
depends on the previous state, are suggested to function as a basis for synaptic
applications.
Spintronics based reservoir computing. Reservoir computing has the goal to
exploit the response of a reservoir to simplify, for example, spatial-temporal recognition tasks. The reservoir itself projects the input into a higher dimensional space,
where it becomes easier to classify. For this concept to work, the reservoir needs to be
a non-linear, complex system with a short-term memory, which is fulfilled by several
physical systems opening up the path for in-materio computing [105]. As spintronics
systems often naturally fulfill these criteria for the reservoir and additionally provide
a lot of tune-ability as well as complexity, together with their low energy consumption, they do provide a promising hardware-based solution for reservoir computing
[106]. It has been proposed that skyrmion fabrics are very well suited for reservoir
computing applications [107].
Stochastic computing. The ansatz of stochastic computing is to trade speed for
accuracy, exploiting the law of large numbers where upon enhancing the number
of experiments the result converges to the expectation value. For example, one can
stochastically multiply two numbers in-between zero and one, when interpreting
them as a probability of having a one in a bit-string. For uncorrelated bit-strings
the multiplication of these two numbers can then be efficiently calculated as sending the two bit-strings through an AND gate. Spintronics offers a potential ansatz
with respect to stochastic computing, as spintronics systems can naturally exhibit
stochastic behavior. Furthermore, recently a device which allows to reshuffle bitstrings based on magnetic skyrmions has been realized [76, 108]. Such a skyrmion
reshuffler allows to restore the decoherence between signals which possibly synchronized. A similar suggestion is to encode the information in probabilistic bits,
also called p-bits. These are bits that fluctuate between 0 and 1 and, in this sense,
interpolate between a classical bit and a q-bit. It has been suggested that magnetic
states naturally provide a realization for such p-bits [109].
Topological quantum computing. Even more exotically, chiral magnetic states
could contribute to topological quantum computing. It has been suggested that Majo-
175
size scale, and manipulability, magnetic textures could play an important role in the
development of such novel computational technologies [98, 99].
Magnetic artificial neural networks. The vast progress within the field of artificial
intelligence is mainly based on the widely enhanced available hardware power, while
most of the concepts have been suggested already a few years ago. So as with deep
artificial neural networks, which nowadays are widely used for different types of AI
applications. However, so far they are mostly performed on the existing hardware
which, due to the classical segmentation in computational units and storage, are
not optimally suited for these types of applications as their power consumption
shows. Instead, alternative architectures which adjust to the deep neural network
structure are proposed, with a focus of creating their central components, i.e., artificial
synapses and neurons, in hardware. There are also several suggestions for magnetic
neuromorphic computing [99, 100] and how to implement artificial neurons [101] and
synapses [102, 103]. In particular, memristors, [104] i.e., devices whose resistance
depends on the previous state, are suggested to function as a basis for synaptic
applications.
Spintronics based reservoir computing. Reservoir computing has the goal to
exploit the response of a reservoir to simplify, for example, spatial-temporal recognition tasks. The reservoir itself projects the input into a higher dimensional space,
where it becomes easier to classify. For this concept to work, the reservoir needs to be
a non-linear, complex system with a short-term memory, which is fulfilled by several
physical systems opening up the path for in-materio computing [105]. As spintronics
systems often naturally fulfill these criteria for the reservoir and additionally provide
a lot of tune-ability as well as complexity, together with their low energy consumption, they do provide a promising hardware-based solution for reservoir computing
[106]. It has been proposed that skyrmion fabrics are very well suited for reservoir
computing applications [107].
Stochastic computing. The ansatz of stochastic computing is to trade speed for
accuracy, exploiting the law of large numbers where upon enhancing the number
of experiments the result converges to the expectation value. For example, one can
stochastically multiply two numbers in-between zero and one, when interpreting
them as a probability of having a one in a bit-string. For uncorrelated bit-strings
the multiplication of these two numbers can then be efficiently calculated as sending the two bit-strings through an AND gate. Spintronics offers a potential ansatz
with respect to stochastic computing, as spintronics systems can naturally exhibit
stochastic behavior. Furthermore, recently a device which allows to reshuffle bitstrings based on magnetic skyrmions has been realized [76, 108]. Such a skyrmion
reshuffler allows to restore the decoherence between signals which possibly synchronized. A similar suggestion is to encode the information in probabilistic bits,
also called p-bits. These are bits that fluctuate between 0 and 1 and, in this sense,
interpolate between a classical bit and a q-bit. It has been suggested that magnetic
states naturally provide a realization for such p-bits [109].
Topological quantum computing. Even more exotically, chiral magnetic states
could contribute to topological quantum computing. It has been suggested that Majo-