Optical Memory and Neural Networks最新文献

Abstract

The nonuniform 2D photovoltaic fields generated near the surface of a photorefractive Fe-doped lithium niobate (LN:Fe) crystal by a nondiffracting optical Bessel beam with concentric ring structures and 532 nm wavelength are used for the trapping of DNA molecules in NaCl buffer on the crystal surface. The simultaneous observation of the long-living Bessel-like refractive lattice recorded in the LN:Fe crystal and the trapped DNA molecules on the crystal surface was performed by an optical phase microscope operating in the transmission mode. With this approach, the DNA molecules are registered as refractive index nonuniformities on the Bessel lattice refractive index pattern. Observations show that DNA molecules are immobilized and trapped at the borderlines of the concentric rings of the refractive lattice recorded by the Bessel beam. The formation of neutral molecular clusters of DNA by Na⁺ counterions with a nearly globular shape and cluster average size of ~4 μm is revealed. A physical model is developed for the analysis of the electric forces map and explanation of the experimental results. The photovoltaic strategy of trapping and manipulation of micro- and nanoparticles on the crystal surface is promising for the elaboration of the lab-on-a-chip devices operating in an autonomous regime with applications in photonics, micro/nano-electronics and biotechnology.

Abstract

The possibility of the maximum-likelihood algorithm-based deep learning of an optical neural network is considered. Using the optimization of thermodynamic parameters of the network, the algorithm can fail when the network undergoes a phase transition caused by changes of network weights in learning. The approach based on Schraudolph–Kamenetsky [1] and Karandashev–Malsagov [2] algorithms is used in computer simulation. Both algorithms allow the free energy of the system on a planar graph to be computed exactly. The restrictions on the number of negative connections are determined that secure the stability of the system, the absence of the phase transition and unrestrained use of the maximum-likelihood algorithm.

Abstract

We present a novel time-dependent two-state model that describes a constant-amplitude level-crossing field configuration, where the frequency detuning varies within a finite interval. A distinctive feature of this configuration is that the resonance crossing always occurs asymmetrically in time, making it an asymmetric version of the second Demkov-Kunike model. The general solution of the problem is expressed in terms of two independent irreducible linear combinations of the Gauss hypergeometric functions. We analyze the asymptotes of the solution in terms of corresponding quasi-energies and calculate the final transition probability in the case when the system starts from the first quasi-energy state.

Abstract

Based on time-dependent Ginzburg-Landau system of equations, Éliashberg’s kinetic equations and finite element modeling, we analyze phonon emission by the phase-slip centers in superconducting filaments. Our results show that in the dissipative regime with these centers, thin superconducting filaments can be effective in originating not only positive but also negative thermal fluxes, i.e., they both generate and absorb phonons. In a stationary oscillatory regime, at a given moment of time, this generation and absorption of phonons reveals itself as positive and negative spectrum of phonons at different spectral ranges. Moreover, at a given spectral range, the emission reverses its sign during the period of oscillation. This fact is associated with the reciprocation of the energy emission and absorption at different spectral intervals during the oscillation period of the phase-slip center. The integral value of energy over the whole spectral range is time-dependent, being positive for some part of the period and negative for the rest of it. Its time integral over the period reveals a positive value, which corresponds to the total energy released in this dissipative state of superconducting filament. In a simple case, when the filament is embedded in a thermal heat bath (substrates typically play that role), this energy dissipates, elevating locally the temperature of filament’s environment. However, in a more sophisticated design, the positive and negative fluxes may become separated. This can be achieved by using the thermal diode effect (the Kapitza boundaries can play the role of such diodes). Such a separation may yield to the net cooling of some part of the filament environment, while the other part will serve as a heat sink. Thus, with an appropriate design of their thermal surroundings, the phase-slip centers can serve as effective solid-state cooling engines. They may be effective for reducing further the cryostat cold finger temperature; for example, from 1 K to sub-K temperatures.

Abstract—

The possibilities of LaF₃–Er³⁺crystal to obtain cascade lasing with CW pumping at 0.52 μm wavelength are considered. The conditions for the formation of inverse populations between Stark levels of neighboring manifolds are determined. It is shown that, at 100 K and CW pump intensity ({{J}_{p}} geqslant 350,,{{text{W}} mathord{left/ {vphantom {{text{W}} {{text{c}}{{{text{m}}}^{2}}}}} right. kern-0em} {{text{c}}{{{text{m}}}^{2}}}}), it is possible to obtain simultaneously laser radiation at 3.21 and 2.88 μm wavelengths.

Abstract

In strong magnetic fields (0.1–6 kG), many atomic lines closely spaced in frequency appear in the absorption spectrum of alkali metal vapors. Due to the small frequency interval between them and the Doppler broadening of the atomic lines, they are overlapped. For spectral separation and study of individual atomic lines, it is necessary to ensure their spectral narrowing. It is shown that this can be done using the saturated absorption method in an atomic vapor contained in a 30 μm-thick cell filled with Rb vapor. All 10 atomic transitions of Rb D₁ line are spectrally very well resolved in the second derivative of the saturated absorption spectrum. Complete resolution of atomic transitions makes this method useful for the determination of a wide range of magnetic fields. The theoretical curves describe the experimental results very well.

Abstract

The paper discusses the representation of the electromagnetic field of an arbitrarily moving charged particle by means of electric field lines. Expressions for the field line equations are derived on the basis of exact Lienar-Wichert field formulas. Parameterization of field lines by means of light signals (dots) emitted at delayed moments of time allows us to avoid the problem of solving the retardation equation. The resulting nonlinear equations are linearized using the Lorentz transformation applied to the emission rate of these light dots in the particle’s rest frame. These linear equations coincide with the Thomas precession equation, which allows us to state that field lines can be thought of as comprised of light dots that were emitted isotropically in the particle’s rest frame at speed (c). The exact solution of the equations is found in the case when the ratio of the trajectory torsion to the product of the trajectory curvature by the Lorentz factor of the particle is a constant value for the trajectory. The class of such fields in particular includes all flat trajectories. Illustrations of field lines are given for two applications of practical interest – the motion of a charged particle in the field of a plane monochromatic linearly polarized wave and for a helical undulator. In addition, it is shown that the developed mathematical apparatus admits consideration of the superluminal motion of the charge. Exact solutions and illustrations of lines for the superluminal motion of a particle along a circle (superluminal synchrotron radiation) are given.

Segmenting nuclei from histopathology images is a crucial step in the early identification and diagnosis of several diseases. Due to the complexity of histopathology images, accurate nucleus segmentation is not a simple operation. However, convolutional neural networks (CNNs) have recently been revealed to be a viable option. The well-known CNN model, namely the U-Net, demonstrated its image segmentation effectiveness in medical field. However, U-Net has several drawbacks, such as information loss after transmission through particular steps. Another significant one is the likelihood of feature mismatches in the encoder and decoder sub-networks in skip connection, which can lead to the fusing of semantically unrelated information and, as a consequence, fuzzy feature maps throughout the learning process. In order to solve these issues, an improved U-Net architecture called Information Added U-Net with Sharp Block (IASB-U-Net) has been proposed for nuclei segmentation from histopathology images. Information is added to the encoder-decoder path in the proposed model after each layer, and sharpening spatial filters are utilized in place of skip connections. The experimental study over a merged dataset demonstrates that the proposed IASB-U-Net produces competitive results when compared to established CNN models such as U-Net, Dense U-Net, SCPP Net, and LiverNet.

Deploying AI models on resource-constrained devices is indeed a challenging task. It requires models to have a small parameter while maintaining high performance. Achieving a balance between model size and performance is essential to ensuring the efficient and effective deployment of AI models in such environments. Knowledge distillation (KD) is an important model compression technique that aims to have a small model learn from a larger model by leveraging the high-performance features of the larger model to enhance the performance of the smaller model, ultimately achieving or surpassing the performance of the larger models. This paper presents a pipeline-based knowledge distillation method that improves model performance through non-linear feature alignment (FA) after the feature extraction stage. We conducted experiments on both single-teacher distillation and multi-teacher distillation and through extensive experimentation, we demonstrated that our method can improve the accuracy of knowledge distillation on the existing KD loss function and further improve the performance of small models.

The research deals with video encoding using a machine learning-based videoframe approximator. The use of neural networks and hierarchical classifiers is considered in the context of this sort of approximator. Using a machine learning-based hierarchical classifier, the approximator switches at each point of a videoframe between elementary approximators from a predefined set of elementary classifiers. Convolutional filters with parametric orthogonal kernels are used as elementary classifiers. An algorithm for optimizing the hierarchical classifier is considered. The algorithm is based on recursive recalculations of the entropy quality index, which provides a good approximation of the encoded-data size. This sort of videoframe approximator is intended for a video codec using nested representations of videoframes. Real video sequences are used in computational experiments. The results indicate that the use of the videoframe approximator with a hierarchical classifier engaging parametric orthogonal kernels enables a noticeable reduction of the size of the encoded-data array.