This PDF file contains the front matter associated with SPIE Proceedings Volume 11469, including the Title Page, Copyright information, and Table of Contents.

Welcome to the Emerging Topics in Artificial Intelligence 2020 conference

We review our results on Ising machines by spatial light modulation. We report on their performance in simulating spin glasses and solving combinatorial optimization problems. We discuss different annealing strategies and recent developments.

We propose a novel implementation of autonomous photonic neural networks based on optically-addressed spatial light modulators (OASLMs). In our approach, the OASLM operates as a spatially non-uniform birefringent waveplate, the retardation of which nonlinearly depends on the incident light intensity. We develop a complete electrical and optical model of the device and investigate the optimal operational characteristics. We study both, feed-forward and recurrent neural networks and demonstrate that OASLMs are promising candidates for the implement of autonomous photonic neural networks with large numbers of neurons and ultra low energy consumption.

Augmented, mixed and virtual Reality (AR/MR) headsets as well as smart glasses have the potential to revolutionize how we work, communicate, travel, learn, teach, shop and get entertained [1],[2]. An MR headset places virtual content into the user’s view of the real world, either via an optical see-through mode (AR/MR) or video-pass-through mode (VR/MR). Today, the return on investment for MR use has been demonstrated widely for the enterprise and defense sectors, but only partially for consumer. In order to meet the high market expectations especially for the upcoming consumer field, several challenges must be addressed, in a variety of fields: optics, display, imaging, sensing, rendering, and MR content. Across each of these fields, artificial intelligence and deep learning techniques can offer important gains in optimization of key performance criteria, allowing for systems to be more performant on more constrained resources.

Neural networks are powerful tools with many possible new applications in nanophotonics. Here, we show how a deep neural network is capable to develop a generalized model of light-matter interactions in both plasmonic and dielectric nanostructures. Using the local geometry as an input, the model infers the internal fields inside the nanostructures from which secondary quantities can be derived such as near-field distributions, far-field patterns and optical cross sections. The neural network successfully captures plasmonic effects and antenna resonances in metals, magneto-electric modes, anapole and Kerker effects in high-index dielectrics, as well as near-field interactions including induced chirality. The neural network is up to five orders faster than conventional simulations, paving the way for real time control and optimization schemes.

In addition to the celebrated numerical techniques, such as Finite-Element and Finite-Difference methods, it is also possible to predict the scattering properties of optical components using artificial neural networks. However, these machine-learning models typically suffer from a simplicity versus accuracy trade-off. In our work, we overcome this trade-off. We train several neural networks with an indirect goal. Instead of training the net to predict scattering, we try to train it the laws of Optics on a more fundamental level. In this way, we can increase the predictive power and robustness while maintaining a high degree of transparency in the system.

The label noise in AI-aided cancer diagnostics has various origins but often poses a challenge to the data analysis. Misclassified samples in the training set can lead to low accuracy of predictions. In this work, we present strategies of reducing the label noise in the context of dermatofluoroscopy (two-photon fluorescence excitation spectroscopy for early diagnosis of malignant melanoma) and support vector machines (SVMs). The experiments performed on real data set composed of 265 pigmented skin lesions confirm the hypothesis of reduced model accuracy in the presence of label noise. Relabeling and especially removing the supporting vector examples from the training set (100 skin lesions) allow for building models of very high predictive accuracy in diagnosing malignant melanoma as shown on independent data set (165 skin lesions). Furthermore, in the limit of very low data quantity, relabeling of supporting vectors and ensembling are shown to yield models that are more robust to label noise.

I will present a Light-sheet fluorescence microscope (LSFM) for fast volumetric imaging during extended periods of time. In this case, the observation arm of the microscope contains an electrically tunable lens (ETL) that is used to shift the focal position of the collection lens. By moving the light sheet plane in synchrony with the ETL, it is possible to scan the full 3D sample, which remains totally static, at high speeds (25 volumes/s) [2]. This system is used to image the spontaneous Ca2+ activity, as reported by GCaMP fluorescence, in 3D of primary neuron cultures in hydrogels. The field of view is of 300µm x 300µm x 1mm. The imaging speeds allows a proper sampling of the propagation of GCaMP signal in the full observation volume [4]. The obtained data is then processed to calculate the connectivity maps in the 3D neuron cultured in hydrogels.

We present a method to generate multiple virtual stains on an image of label-free tissue using a single deep neural network according to a user-defined micro-structure map. The input to this network comes from two sources: (i) autofluorescence microscopy images of the unlabeled tissue, (ii) a user-defined digital-staining-matrix. This digital-staining-matrix indicates which stain is to be virtually-generated for each pixel, and can be used to create a micro-structured stain map, or virtually blend stains together. We experimentally validated this approach using blind-testing of label-free kidney tissue sections, and successfully generated combinations of H and E, Masson’s Trichome stain, and Jones silver stain.

We present a deep learning-based, high-throughput, accurate colorization framework for holographic imaging systems. Using a conditional generative adversarial network (GAN), this method can be used to remove the missing-phase-related spatial artifacts using a single hologram. When compared to the absorbance spectrum estimation method, which is the current state-of-the art method used to perform color holographic reconstruction, this framework is able to achieve a similar performance while requiring 4-fold fewer input images and 8-fold less imaging and processing time. The presented method can effectively increase the throughput for color holographic microscopy, providing the possibility for histopathology in resource limited environment.

Particles with dimensions smaller than the wavelength of visible light are essential in many fields. As particle size and composition greatly influence particle function, fast and accurate characterisation of these properties is important. Traditional approaches use the Brownian motion of the particles to deduce their size, and therefore requires to observe the particles for many consecutive time-steps. In addition, such techniques can only be applied in environments with known viscosity, hindering characterization in complex environments. In this work, we demonstrate characterisation of subwavelength particle size and refractive index surpassing that of traditional methods using two orders of magnitude fewer observations of each particle, with no reference to particle motion. This opens up the possibility to characterise and temporally resolve the properties of subwavelength particles in complex environments where the relation between particle dynamics and size is unknown.

Emerging deep learning based computational imaging techniques promise unprecedented imaging capabilities. In this talk, I will discuss our efforts in building such techniques with improved reliability and explainability. First, I will discuss a physics-guided deep learning imaging approach that enables designing highly efficient multiplexed data acquisition schemes for intensity diffraction tomography. I will highlight an uncertainty quantification framework to assess the reliability of the deep learning predictions. Second, I will present a deep learning approach to invert the effects of scattering. The trained network is able to make high-quality object predictions from speckles captured from diffusers entirely different from the training data and is highly robust to defocus across several depths of field. I will highlight a dimensionality reduction approach to explain the underlying statistics learned by the network and understand the capability and limitations for generalization.

In-memory signal processing and computing based on the integrated phase-change photonic platform

Maximal computing performance can only be achieved if neural networks are fully hardware implemented. Besides the potentially large benefits, such parallel and analogue hardware platforms face new, fundamental challenges. An important concern is that such systems might ultimately succumb to the detrimental impact of noise. We study of noise propagation through deep neural networks with various neuron nonlinearities and trained via back-propagation for image recognition and time-series prediction. We consider correlated and uncorrelated, multiplicative and additive noise and use noise amplitudes extracted from a physical experiment. The developed analytical framework is of great relevance for future hardware neural networks. It allows predicting the noise level at the system’s output based on the properties of its constituents. As such it is an essential tool for future hardware neural network engineering and performance estimation.

Performing feature extractions in convolution neural networks for deep-learning tasks is computational expensive in electronics. Fourier optics allows convolutional filtering via dot-product multiplication in the Fourier domain similar to the distributive law in mathematics. Here we experimentally demonstrate convolutional filtering exploiting massive parallelism (10^6 channels, 8-bit at 1kHz) of digital mirror display technology, thus enabling 250 TMAC/s. An FPGA-PCIe board controls the ‘weights’ and handles the data I/O, whereas a high-speed camera detects the inverse-Fourier transformed (2nd lens) data. Gen-1 processes with a total delay (including I/O) of ~1ms, while Gen-2 at 1^-10ns leveraging integrated photonics at 10GHz and changing the front-end I/O to a joint-transform-correlator (JTC). These processors are suited for image/pattern recognition, super resolution for geolocalization, or real-time processing in autonomous vehicles or military decision making.

This presentation will discuss demonstrations of real-time operation of photonic neural networks going beyond machine learning applications. It contains a review of how a wavelength-based neuron can be implemented using silicon photonics. We discuss three application examples: principal component analysis for blind source separation, fiber nonlinearity compensation in long-haul communication systems and ODE system solving in real-time.

Computational imaging involves the joint design of imaging system hardware and software, optimizing across the entire pipeline from acquisition to reconstruction. Computers can replace bulky and expensive optics by solving computational inverse problems. This talk will describe end-to-end learning for development of new microscopes that use computational imaging to enable 3D fluorescence and phase measurement. Traditional model-based image reconstruction algorithms are based on large-scale nonlinear non-convex optimization; we combine these with unrolled neural networks to learn both the image reconstruction algorithm and the optimized data capture strategy.

We present a super-resolution framework for coherent imaging systems using a generative adversarial network. This framework requires a single low-resolution input image, and in a single feed-forward step it performs resolution enhancement. To validate its efficacy, both a lensfree holographic imaging system with a pixel-limited resolution and a lens-based holographic imaging system with diffraction-limited resolution were used. We demonstrated that for both the pixel-limited and diffraction-limited coherent imaging systems, our method was able to effectively enhance the image resolution of the tested biological samples. This data-driven super resolution framework is broadly applicable to various coherent imaging systems.

We present a deep learning-based framework to perform single image super-resolution of SEM images. We experimentally demonstrated that this network can enhance the resolution of SEM images by two-fold, allowing for a reduction of the scanning time and electron dosage by four-fold without any significant loss of image quality. Using blindly tested regions of a gold-on-carbon resolution test target, we quantitatively and qualitatively confirmed the image enhancement achieved by the trained network. We believe that this technique has the potential to improve the SEM imaging process, particularly in cases where imaging throughput and minimizing beam damage are of utmost importance.

Here we introduced an integrated photonics-based TPU by strategically utilizing a) photonic parallelism via wavelength division multiplexing, b) high 2 Peta-operations-per second throughputs enabled by 10’s of picosecond-short delays from optoelectronics and compact photonic integrated circuitry, and c) zero power-consuming novel photonic multi-state memories based on phase-change materials featuring vanishing losses in the amorphous state. Combining these physical synergies of material, function, and system, we show that the performance of this 8-bit photonic TPU can be 2-3 orders higher compared to an electrical TPU whilst featuring similar chip areas. The runtime complexity is O(1) once the kernel matrix is programmed, however the engine scales with O(N^3) resources (devices). This system could ultimately perform in the range 10-500fJ/MAC, 1-50 TMACs/mm^2, and ~100ps (1 clock cycle) per VMM operation.

We present a diffractive deep neural network-based framework that can simultaneously process a continuum of illumination wavelengths to perform a specific task that it is trained for. Based on this framework, we designed and 3D printed a series of optical systems including single and double pass-band filters as well as a spatially-controlled wavelength de-multiplexing system using a broadband THz pulse as input, revealing an excellent match between our numerical design and experimental results. The presented optical design framework based on diffractive neural networks can be adapted to other parts of the spectrum and be extended to create task-specific metasurface designs.

The term "computing" has a specific, well-defined, powerful, traditional meaning -- condensed in the paradigm of Turing computability (TC). A core aspect of TC is the perfectly reliable composition of perfectly identifiable symbolic tokens into complex, hierarchical symbolic structures. But all which is novel and promising and original in "neuromorphic" information processing leads away from such perfect symbolic compositionality. Apparently new formal conceptions of "computing" would be most welcome (and a new term for it, too). In this talk I will explain the principle of Turing computability for a non-CS audience, and then proceed to carve out a number of concrete examples of „computational“ phenomena that separate neuromorphic information processing from Turing computability (and hence, from all known digital computing). Some of these items are classical topics in the philosophy of AI, others having more recently emerged from technological progress in non-digital hardware. I conclude with a proposal for a particular search direction for exploring a new kind of formal theory which might give the field of neuromorphic computing a unified foundation, similar in power and beauty to Turing computability in the field of digital computing.

Structural connectomics uses tractography for inferring the physical connections between brain regions. An issue of current tractography methods is their large false-positive rate, which makes them not reliable enough for assessing structural connectivity. An approach to deal with this problem is tractogram filtering in which anatomically implausible streamlines are discarded as a post-processing step after tractography. In this talk, I will first review the main approaches and methods that have been proposed in the literature for tractogram filtering and their limitations. Second, I will describe our efforts of using artificial intelligence for improving the accuracy of structural connectivity analyses. Finally, I will give a perspective on the main challenges for the development of new methods in this field in the next few years.

Computed tomography (CT) is a widely available, low-cost neuroimaging modality primarily used as a brain examination tool for visual assessment of structural brain integrity in neurodegenerative diseases such as dementia disorders. In this study, we developed a deep learning model to expand the applications of CT to morphological brain segmentation and volumetric extraction. We trained densely connected 3D convoluted neural network variants called U-Nets to segment intracranial volume (ICV), grey matter (GM), white matter (WM) and cerebrospinal fluid (CSF). Dice similarity scores and volumetric Pearson correlation were the evaluation metrics incorporated. Our pilot study created a model that enables automated segmentation in CT with results comparable to magnetic resonance imaging.

Mesial temporal sclerosis (MTS) is the principal cause of complex epilepsy, is manifested principally by gliosis and hippocampal volume loss. This project aims to develop an algorithm that allows automatic measurement of hippocampal volume and signal intensity in magnetic resonance imaging. The algorithm developed uses preprocessing of the images to reduce the artifacts and for the extraction of the features were used techniques of machine learning (support vector machine) and texture analysis. Results can help to optimize time in the assessment of the mesial temporal sclerosis and can contribute to the best training to the youngers neuroradiologists.

Parkinson's disease (PD) is a complex neurodegenerative disorder characterized by motor and non-motor deficits. Several studies found changes in the topological organization of functional networks in PD built by methods that assume a simultaneous and undirected activation between brain areas. However, changes associated with PD may result in a specific alteration of the directed activity patterns between brain areas. In this study, we propose a new method to build directed functional networks in patients with PD. We show that the directed network analyses can identify widespread functional brain changes in PD characterized by higher efficiency, clustering and transitivity as well as lower modularity. Some of these network measures were associated with motor, executive and memory deficits, suggesting they are sensitive to clinical impairment in PD. Altogether our findings suggest that the directional flow in brain activation could be used as an indicator of PD-related neuronal changes.

The failure of machine learning models to accurately capture causal relationships is becoming increasingly well understood, leading to an explosion of casually motivated machine learning techniques. In this talk we consider the quantum case: can we learn causal structure in a quantum world? We introduce a quantum causal machine learning framework and also consider the possibility of a quantum enhanced agent that learns via explicitly quantum interventions. The latter model of a quantum agent uses optical probes to learn about the external world in a manner that improves upon the classical analogue.

Artificial microswimmers are designed to mimic the self-propulsion of microscopic living organisms to yield access to the complex behavior of active matter. As compared to their living counterparts, they have only limited ability to adapt to environmental signals or retain a physical memory. Yet, different from macroscopic living systems and robots, both microscopic living organisms and artificial microswimmers are subject to thermal noise as a key feature in microscopic systems. Here we combine real-world artificial active particles with machine learning algorithms to explore their adaptive behavior in a noisy environment with reinforcement learning. We use a real-time control of self-thermophoretic active particles to demonstrate the solution of a standard navigation problem with single and multiple swimmers and show that noise decreases the learning speed, increases the decision strength and modifies the optimal behavior based on a delayed response in the noisy environment.

The next generation of multi-functional optical materials will customize electric field responses via a careful arrangement of micro- and nano- scale scatterers to achieve targeted optical performance otherwise unattainable in traditional bulk media. Macroscopically, such designed materials collectively respond to radiation based on the geometric shape, distribution, and inherent material properties of these sub-wavelength structures. The core challenge is in prescribing a configuration which results in a desired property. It becomes immediately clear these metamaterial systems pose significant challenges because of the near-infinite design space one needs to consider. Recently, artificial neural networks (ANNs) have been used to successfully approach this intractable problem. In the specific context of designing an all-dielectric metasurface reflector, we showed that joining two properly trained ANNs can both emulate Maxwell equations as well as inversely correlate reflection and transmission spectra.¹ Though highly accurate ANNs were trained, the ANNs employed were never optimized in terms of architecture (e.g. number of layers, number of nodes, shape of the network) or hyperparameters (e.g. batch sizes, activation functions, loss functions). In this manuscript, we apply Bayesian optimization with Gaussian processes to first optimize the architecture and then the hyperparameters of the spectra predicting networks described in Ref. 1. The goal is not only to improve upon the previously implemented ANNs but to analyze the effect of different ANN architectures and convergence settings on overall spectra predictive performance.

Machine learning (ML) has emerged in recent years as a data-driven approach for photonic inverse design. Despite their impressive performance in finding abstract mappings between the design parameters and optical properties, ML algorithms suffer from a high likelihood of slow converging when there exist multiple designs giving similar optical responses. Here we adopt a deep convolutional mixture density neural network, which models the design as a mixture of Gaussian distributions rather than discrete values, to address the non-uniqueness issue. An example of layered structures consisting of alternating oxides under arbitrary incidence conditions is present to showcase the proof of concept.

In this study, we present an intelligent inverse design method of nanophotonic structures using a data-driven approach of deep learning [1-3]. We demonstrate attempts to increase the degree of freedom of design possibility, by designing arbitrary shapes of nanophotonic structures [2] and by designing types of materials and design parameters at the same time [3]. Recently, deep learning has shown its capability to provide appropriate nanophotonic designs for given desired optical functionalities. So far, however, these approaches have been applied to design only few structural parameters for the pre-defined nanophotonic systems. Here, we show two inverse design of nanophotonic structures using deep learning. Our methods increase the degree of freedom by designing nanophotonic structures in the form of binary images and designing types of materials and structural parameters simultaneously.

The calibration of physical force fields from particle trajectories is important for experiments in soft matter, biophysics, active matter, and colloidal science. However, it is not always possible to have a standard method to characterize a force field, especially for systems that are out of equilibrium. Here, we introduce a generic toolbox for calibrating any kind of conservative or non-conservative, fixed or time-varying potentials that is powered by recurrent neural networks (RNN). We show that with the help of neural networks, we can outperform standard methods as well as analyze systems that cannot be approached by existing methods. We provide a software package that is available online for free access.

Computationally accurate methods for simulating optical tweezers tend to be prohibitively slow, limiting their use to only very simple problems. Simplified models, such as the harmonic model, enable larger simulations by trading off accuracy for speed. In this presentation, I will demonstrate how training an artificial neural network to predict optical force combines the speed of a harmonic model with the accuracy of a semi-analytical method. Artificial neural networks not only enable more extensive and accurate dynamics simulations, but also collaboration through sharing of pre-trained models which can easily be distributed and used on mobile devices and in web browsers, as I will demonstrate.

From the start of digital video microscopy over 20 years ago, single particle tracking has been dominated by algorithmic approaches. These methods are successful at tracking well-defined particles in good imaging conditions but their performance degrades severely in more challenging conditions. To overcome the limitations of traditional algorithmic approaches, data-driven methods using deep learning have been introduced. They managed to successfully track colloidal particles as well as non-spherical biological objects, even in unsteady imaging conditions.

Holographic microscopy encodes the 3D information of a sample into a single hologram. However, holographic images are in general inferior to bright-field microscopy images in terms of contrast and signal-to-noise ratio, due to twin-image artifacts, speckle and out-of-plane interference. The contrast and noise problem of holography can be mitigated using iterative algorithms, but at the cost of additional measurements and time. Here, we present a deep-learning-based cross-modality imaging method to reconstruct a single hologram into volumetric images of a sample with bright-field contrast and SNR, merging the snapshot 3D imaging capability of holography with the image quality of bright-field microscopy.

We report new design strategies to increase the inference accuracy of Diffractive Deep Neural Networks (D2NNs). Using a differential detection scheme that is combined with the joint-training of multiple D2NNs, each specialized on a single object-class, D2NN-based all-optical classification systems numerically achieve blind-testing accuracies of 98.52%, 91.48% and 50.82% for MNIST, Fashion-MNIST and grayscale CIFAR-10 datasets, respectively. Furthermore, using three independently-trained D2NNs that project their light onto a common output plane enables the system to achieve 98.59%, 91.06% and 51.44%, respectively. Through these systematic improvements, the reported blind-inference performance sets the state-of-the-art for an all-optical neural network design.

DeepTrack is an all-in-one deep learning framework for digital microscopy, attempting to bridge the gap between state of the art deep learning solutions and end-users. It provides tools for designing samples, simulating optical systems, training deep learning networks, and analyzing experimental data. Moreover, the framework is packaged with an easy-to-use graphical user interface, designed to solve standard microscopy problems with no required programming experience. By specifically designing the framework with modularity and extendability in mind, we allow new methods to easily be implemented and combined with previous applications.

The deviation from pure Brownian motion, generally referred to as anomalous diffusion, has received large attention in the scientific literature to describe many physical scenarios. Several methods, based on classical statistics and machine learning approaches, have been developed to characterize anomalous diffusion from experimental data, which are usually acquired as particle trajectories. With the aim to assess and compare the available methods to characterize anomalous diffusion, we have organized the Anomalous Diffusion (AnDi) Challenge (http://www.andi-challenge.org/). Specifically, the AnDi Challenge will address three different aspects of anomalous diffusion characterization, namely: (i) Inference of the anomalous diffusion exponent. (ii) Identification of the underlying diffusion model. (iii) Segmentation of trajectories. Each problem includes sub-tasks for different number of dimensions (1D, 2D and 3D). In order to compare the various methods, we have developed a dedicated open-source framework for the simulation of the anomalous diffusion trajectories that are used for the training and test datasets. The challenge was launched on March 1, 2020, and consists of three phases. Currently, the participation to the first phase is open. Submissions will be automatically evaluated and the performance of the top-scoring methods will be thoroughly analyzed and compared in an upcoming article.

Biology has changed radically in the last two decades, transitioning from a descriptive science into a design science. Synthetic biology leverages engineering approaches to produce biological systems to a given specification (e.g. produce x grams of a medical drug or invade this type of cancer cell). In this effort, new tools are now available that promise to disrupt this field: from CRISPR-enabled genetic editing, to high-throughput omics phenotyping, and exponentially growing DNA synthesis capabilities. However, our inability to predict the behavior of bioengineered systems hampers synthetic biology from reaching its full potential. We will show how the combination of machine learning and automation enables the creation of a predictive synthetic biology for the benefit of society.

We present an adapted single shot neural network architecture (YOLO) for the real-time localization and classification of particles in optical microscopy. Our work is aimed at the manipulation of microscopic objects in real-time by a feedback loop. The network is implemented in Python/Keras using the TensorFlow backend. The trained model is then exported to a GPU supported C library for real-time inference readily integrable in other programming languages such as C++ and LabVIEW. It is capable of localizing and classifying several hundred of microscopic objects even at very low signal-to-noise ratios running for images as large as 416 x 416 pixels with an inference time of about 10 ms. We demonstrate real-time detection in tracking and manipulating active particles of different types. Symmetric active particles, as well as Janus particles propelled by self-thermophoretic laser-induced processes, are identified and controlled via a Photon-Nudging procedure developed in the group.

We present a high-throughput and cost-effective computational cytometer for rare cell detection, where the target cells are specifically labeled with magnetic particles and exhibit an oscillatory motion under a periodically-changing magnetic field. The time-varying diffraction patterns of the oscillating cells are then captured with a holographic imaging system and are further classified by a customized pseudo-3D convolutional network. To evaluate the performance of our technique, we detected serially-diluted MCF7 cancer cells that were spiked in whole blood, achieving a limit of detection (LoD) of 10 cells per 1 mL of whole blood.

We report a point-of-care (POC) assay and neural network-based diagnostic algorithm for Lyme Disease (LD). A paper-based test in a vertical flow format detects 16 different IgM and IgG LD-specific antibodies in serum using a mobile phone reader and automated image processing to quantify its colorimetric signals. The multiplexed information is then input into a trained neural-network which infers a positive or negative result for LD. The assay and diagnostic decision algorithm were validated through fully-blinded testing of human serum samples yielding an area-under-the-curve (AUC), sensitivity, and specificity of 0.950, 90.5%, and 87.0% respectively, outperforming previous Lyme POC tests.

We demonstrate a deep learning-based technique which digitally stains label-free tissue sections imaged by a holographic microscope. Our trained deep neural network can use quantitative phase microscopy images to generate images equivalent to the same field of view of the specimen, once stained and imaged by a brightfield microscope. We prove the efficacy of this technique by implementing it with different tissue-stain combinations involving human skin, kidney, and liver tissue, stained with Hematoxylin and Eosin, Jones’ stain, and Masson’s trichrome stain, respectively, generating images with equivalent quality to the brightfield microscopy images of the histochemically stained corresponding specimen.

The existing neural networks activation functions, the Sigmoid among others is the only set uses across various applications of NNs such as microscopy, neuromorphic, optical, robotics, finance and transportation. Only one set applies to different areas of application. Also, these activation functions’ selections are based on trial and error, neither emanate from the AI and or training datasets, nor from the testing data. This formed NNs’ Black-box. Jamilu (2019) proposed that strong links between the AI and or training datasets and activation functions must be established. This is to replace the NNs’ Black-box models with the models rely much less on experts’ assumptions, and much more on input AI and or training datasets, time change and specific area of application. Thus, Jamilu (2019) proposed Criterion(s) for the rational selection of activation functions. The paper is to use superintelligent NNs for stock price predictions, portfolio optimization, and general application approaches to shed light on the paper’ title.

Currently, we are dealing with a very limited set of activation functions such as Sigmoid, ReLu, Leaky ReLu among others. These activation functions used in the existing digital brain neural network systems are chosen using assumption with the help of “trial and error” approach. However, they do not ethically and appropriately establish any relationship with the referenced AI datasets. Jamilu (2019) proposed that a digital brain should have at least 2000 to 100 billion distinct activation functions implying distinct artificial neurons satisfies Jameel’s criterion(s) for it to normally mimic the human brain. The objectives of this paper are to propose a theorem called “Digital Brain Completeness Theorem”, “superintelligent digital brain neural network systems” and why it is tremendously important to have an extremely huge distinct activation functions implying distinct artificial neurons in a digital brain just like in the case of its counterpart for it to function rationally.

We report a deep learning-based framework which can be used to screen thin blood smears for sickle-cell-disease using images captured by a smartphone-based microscope. This framework first uses a deep neural network to enhance and standardize the smartphone images to the quality of a diagnostic level benchtop microscope, and a second deep neural network performs cell segmentation. We experimentally demonstrated that this technique can achieve 98% accuracy with an area-under-the-curve (AUC) of 0.998 on a blindly tested dataset made up of thin blood smears coming from 96 patients, of which 32 had been diagnosed with sickle cell disease.

Currency classification is an important task in computer vision. Traditional models extract relevant features (brightness, shape, colour etc.) through complex mathematical calculations. A deep learning approach towards fiscal classification by automatically learning higher order feature representations of the dataset is presented. A family of Resnet models is trained to minimize the effect of distortions in currency dataset. The classifier achieves a peak test set performance of 98.09% and an ensembled accuracy of 98.3%. Finally, an optimization method is introduced to allow the models initialized with pretrained weights to converge faster and achieve better accuracy in certain cases.

Digital Holographic Microscopy (DHM) has been a successful imaging technique for various applications in biomedical imaging, particle analysis, and optical engineering. Though DHM has been successful in reconstructing 3D volumes with stationary objects, it has still been a challenging task to track fast mobile objects. Recent advancements in deep learning with convolutional neural networks have been proven useful in solving experimental difficulties, starting from tracking single particles to multiple bacterial cells. Here, we propose a compact DHM driven by neural networks with a minimal amount of optical elements with an ultimate aim for easy usage and transportation.

This work uses a deep learning approach using convolutional neural networks to locate and classify nanostructures in a heterogenous composition material from TEM imaging. We developed a methodology that allowed us to create 533 ground truth of TEM images with three different classes: 1) silicon oxide nanoparticles, 2) yttrium silicate particles and 3) silicon oxide coating. We performed the classification, location, and segmentation of chemical compounds reaching scores above 80% of accuracy using Mask R-CNN architecture with Anaconda Python 3.7 and the Tensorflow framework under Windows 10.

The brain is a complex network that relies on the interaction between its various regions, known as the connectome. The organization of the human connectome has been studied on different imaging modalities using a single network approach. Here, we integrate the networks obtained from amyloid positron emission tomography (PET) and structural magnetic resonance imaging (MRI) data into multilayer networks using BRAPH 2.0 (BRain Analysis using graPH theory, http://braph.org/) and compare these networks between patients with Alzheimer’s Disease and controls. Multilayer modularity, multi-participation coefficient, and multilayer motifs are calculated, and group comparisons are carried out using permutation testing. The study of multilayer brain networks is a promising new field that can potentially provide new insights into the interaction between anatomical, functional and metabolic brain connectivity.

Recently, deep neural network (DNN) based adaptive optics systems were proposed to address the issue of latency in existing wavefront sensorless (WFS-less) aberration correction techniques. Intensity images alone are sufficient for the DNN model to compute the necessary wavefront correction, removing the need for iterative processes and allowing practical real-time aberration correction to be implemented. Specifically, we generate the desired aberration correction phase profiles utilizing a DNN based system that outputs a set of coefficients for 27 terms of Zernike polynomials. We present an experimental realization of this technique using a spatial light modulator (SLM) on real physical turbulence-induced aberration. We report an aberration correction rate of 20 frames per second in this laboratory setting, accelerated by parallelization on a graphics processing unit. There are a number of issues associated with the practical implementation of such techniques, which we highlight and address in this paper.

The brain connectome can be modeled as a large-scale complex network characterized by high clustering and short network paths. Most studies assess these properties by comparing them to a null network model, generated by randomly rewiring the edges between the nodes of the original network, known as edge swapping. However, this method is computationally expensive and time consuming, mainly in networks with a high number of connections. In this study, we developed an alternative method to create null network models, the allin method. We show that both methods compute null networks with comparable topology, however, the allin method performed the randomization procedure in noticeably less time. The allin method is particularly more effective in the case of high-resolution networks and relatively higher densities. As such, these results suggest that the allin method is a more time efficient alternative to compute null model networks compared to the traditionally used edgeswap method.

There is increasing evidence showing that graph theory is a promising tool to study the human brain connectome. By representing brain regions and their connections as nodes and edges, it allows assessing properties that reflect how well brain networks are organized and how they become disrupted in neurological diseases such as Alzheimer’s disease, Parkinson’s disease, epilepsy, schizophrenia, multiple sclerosis and autism. Here, we present BRAPH 2.0 (BRain Analysis using graPH theory version 2.0), which is a major update of the first object-oriented open source software written in Matlab for graph-theoretical analysis that also implements a graphical interface (GUI). BRAPH utilizes the capability of object-oriented programming paradigm to provide clear, robust, clean, modular, maintainable, and testable code.

This paper presses a developed methodology of estimating the total number of charging points in the Electric Vehicle Charging Station (EVCS). Three various EVCSs in the urban core, suburban area and the rural area were modeled and investigated by using an established database for fourteen different Electric Vehicles (EVs) of different manufacturers. Monte-Carlo simulation technique (MCST) was applied with high-dense iterative runs to predict the peak hour energy demand that can be occurred in the proposed three zones besides expecting the arrival interval time of the EVs across the day according to the percentage of daily demand of each station. Moreover, an imperially constructed equation is used to calculate the number of charging points in each zone by estimating the normalized arrival time with the aid of MCST. The precise estimating of the total number of charging points for each station is minimizing the charging time and the queuing delay issues.

Syntronic AB collaborates with the Karolinska Institute and the University of Gothenburg in developing the software BRAPH (Brain Analysis with graPH theory). BRAPH is used for analyzing brain graphs derived from brain imaging data. With the increasing prevalence of neurodegenerative disorders, there is a need for new reliable methods for diagnosis to help implementing preventive treatments before damage is widespread. To this aim, Syntronic has developed a new Python version of BRAPH and investigated the potential use of implementing artificial intelligence to the software. In particular, the ability of different classes of Artificial Neural Networks (ANN) in detecting Alzheimer’s disease from MRI derived brain graphs has been studied. The study showed promising results using Graph Convolutional Neural Networks; a class of ANNs that generalize convolutions to graphs. The goal is to implement a new machine learning toolbox to BRAPH to exploit the benefits of these algorithms.

Identifying a particle in an optical trap can be a difficult task, especially for biological samples with low contrast. The relationship of radius and refractive index to the stiffness of optical traps is non-intuitive, motivating a machine learning approach. We demonstrate methods for real-time estimates of the radius and refractive index of particles trapped by optical tweezers. This is achieved by analysing the particle’s position and force with artificial neural networks. Our network achieved binary classification of experimental particles by sampling only milliseconds of force and position values. This demonstrates that real-time particle recognition is achievable with machine learning systems.