Hardware implementation of a neuron for neuromorphic optical system

. Specific features of neuromorphic systems construction and aspects of basic elements functions realisation are discussed. Hardware implementations of basic elements of pulsed neuromorphic optical system - synapses and neurones - based on the phase state change of thin-film structures made of chalcogenide materials under the influence of laser pulses are described. Calculations of optical characteristics of GeSbTe thin films in multilayer structures have been carried out. The hardware implementation of the main functions of optical neurons was analysed on the model with multilayer coatings based on phase-changeable materials.


Introduction
Neuromorphic technology refers to computing systems that are inspired by the biological architectures of the nervous system.Neuromorphic systems are widely used for pattern recognition and classification with many potential applications from fingerprint, iris and face recognition to target detection etc.The parameters of neuromorphic systems are generated by training on a dataset, as a result, the neural network is able to continuously recognise or classify the same type of data in real time.Currently, neuromorphic systems and networks are usually implemented as software algorithms with hardware implementation using technologically proven CMOS technology.Multiprocessor parallel computing on graphics graphics cards is used for training, and end devices ("inference", "edge") allow to run trained networks with low power consumption.There are known methods of software implementation of formal neural systems [1,2] with training on large classified data sets for a certain task and object, and architecture of direct passage of the processed signal to apply the method of error back propagation.A large number of variants of formal neural systems on elements with changing electrophysical properties have been implemented [3,4,5].Execution of neural network algorithms by these methods requires orders of magnitude more energy for their work in comparison with the biological brain of a mammal.
Formal neural networks have significant disadvantages: a) training is possible only on large marked data for each narrowly specific task, b) the architecture is limited by the requirement of direct signal passage for the possibility of training by the method of error back propagation, c) each training cycle (recording of weights) requires energy for all synapses, d) the neuron implements the summation function and with a normalised output constant signal.
The consequence of the above disadvantages is that formal neural networks perform a sequential process of training and recognition with a lack of perceptual integrity when working in a fixed feature space that is the same for all recognition objects.Further expansion of the functional by type of tasks leads to a decrease in efficiency and performance, and raises the problem of searching for other hardware implementations of neural network algorithms.

Basic elements of neuromorphic systems: synapses and neurons
Current hardware implementations of deep neural networks are still far from competing with biological neural systems that demonstrate the benefits of fault tolerance and energy efficiency for solving real-world problems involving visual information, sound recognition, and motion control.Humans can easily recognise different objects and make sense of large amounts of visual information in complex real-world environments using sophisticated computations that are more robust, malleable and fault-tolerant than any current digital computer.
Artificial neural system models are much simpler, but they offer a general computational framework for a wide range of tasks.An artificial neuron realises two functions: integration and non-linear activation.Integration involves input signals -synapses.Synapses are usually much larger than neurons (approximately ten thousand synapses per neuron in the case of the human brain).The activation function is applied to a weighted signal.At the output, a neuron can generate a continuous signal or pulses ("action potentials", "bursts", "spikes") when the integrated input signal exceeds a threshold value ("membrane potential") [6].The output of a neuron is translated or branched to many other neurons via synapses.Synapses change the strength of their connection due to neuronal activity, and updating the weight of this connection is called network learning.A key challenge in neuromorphic computing technology is to develop compact devices that mimic the functions of biological synapses and neurons.
One variant of an array of synapses and neurons is implemented in the IBM TrueNorth neuromorphic processor [7], where the weight coefficients of the trained network are written in integers in a single artificial synapse from an array of semiconductor transistors, and the neuron generates pulses using the LIF (Leaky Integrate and Fire) model.Another variant of a pulse system with on-chip learning is presented in Intel's developments on the Loihi processor [8].
Pulse neural networks use local learning rules, updating weights only for activated neurons and their immediate neighbourhood, forming more promising computing power while minimizing the amount of marked data [9].The most popular and implemented in artificial analogues are local learning rules based on the STDP (Spike-Timing Dependent Plasticity) method, using the approaches of biological neural networks [7].

Neuromorphic optical systems
Architectural concepts and implementations of various neuron models, learning methods and topologies are already being actively researched.This diversity implies that research in neuromorphic computing should not be expected to lead to a single successful implementation or a single application.Ongoing research is needed to identify applications where photonics will be most superior to the ever-improving state-of-the-art in electronic computing.
Energy efficiency, noise immunity, speed, reliability, and scalability are key characteristics of current neuromorphic system architectures.The basic elements of such systems should provide minimum loss, contrast and signal stability; high switching speed with simple design based on proven technologies.The advanced development of neuromorphic systems is possible due to the transition to pulse neural networks realised with the help of optical technologies.The development of such systems is possible through the development of integrated optical and electrical devices or new architectures with laser emitters.Our group is developing a pulsed neuromorphic optical system (see Fig. 1), which contains optical synapses 1, optical neurons 2, where in one nucleus 8 the synapses are influenced by pulses of laser radiation in the visible range 5, thus changing the intensity of pulses of laser radiation in the near infrared range 6.Further, the pulses 6 are amplified in the amplifierintegrator 3 and jointly act on the neuron, whereby the pulse from the neuron is divided in the amplifier-splitter 4 into the same intensity output pulses of near-infrared range 7, which are input signals for the following nuclei of the neuromorphic optical system.The optical synapse and optical neuron are realised on the basis of planar waveguides and multilayer structures with thin lumenal films of phase-changeable materials.Chalcogenides, selenides, sulfides, tellurides, antimonides, and transition metal oxides are used as phase-changeable materials.For chalcogenides, due to high extinction in the wavelength range of electromagnetic radiation from 400 to 800 nm, laser induction of thin films is possible with pulsed sources.The wavelength of modulated radiation in the near infrared is due, firstly, to reliable technological solutions based on planar and solid-state amplifiers, and, secondly, to the high contrast of optical properties of the amorphous and crystalline phase of chalcogenide thin films.
The amplifier-integrator and amplifier-splitter are made in the form of solid-state elements doped with rare-earth materials, with a given geometry and spatial distribution of the optical pump radiation.

Aspects of hardware implementation of an optical neuron
The proposed model of hardware implementation of the optical neuron is based on a thinfilm structure with layers of phase-changeable materials supplemented with enlightening coatings (see Fig. 2).The measured optical characteristics of the obtained structures differ for two thicknesses (30 and 60 nm) of GeSbTe films and their phase state in accordance with the model.To estimate the contrast of optical characteristics, it is convenient to represent delta (1) of the transmittance coefficients when the phase state of the GeSbTe film changes from the amorphous state to the crystalline state (see Fig. 3).

∆𝑇𝑇/𝑇𝑇=
−    (1) For such structures, the transmittance delta decreases upon the phase transition of both films for the probe laser, while for the "O-probe" (1260 nm) and "C-probe" (1550 nm) the transmittance decreases, but the contrast for optical characteristics is minimal.

Si
When analysing the multilayer structure of an optical neuron, attention should be paid to the significant degradation of the properties of the layers of the structure, which leads to the impossibility of using switching between amorphous and crystalline phases as a principle of optical neuron operation.To increase the switching cycles between different states, it is necessary to use materials that have two or more crystalline phases with optical property contrast.It is also necessary to exclude thermal processes that lead to the appearance of amorphous phase in the material during switching.

Conclusion
The prospects of creating a pulsed neuromorphic optical system are related to the development of the following directions: research and synthesis of new functional materials and methods of changing their phase state, development of new methods of training artificial neural systems, transition to the technologies of three-dimensional "on-chip" systems [10,11,12], creation of compact and highly efficient solid-state pulsed laser emitters.Special attention should be paid to the creation of three-dimensional photonic systems from the integration of optoelectronic elements to external interfaces and data transmission lines.Maximum integration of components into controllers and processors is required to move from electrical to optical interfaces.This requires heterogeneous laser structures, photovoltaic converters and optical signal processing units mounted in integrated circuits.All of this is now more a matter of basic and applied science than of production.

Fig. 2 .
Fig. 2. Model of hardware implementation of optical neuron based on GST.