skip to main content

Title: Freely scalable and reconfigurable optical hardware for deep learning

As deep neural network (DNN) models grow ever-larger, they can achieve higher accuracy and solve more complex problems. This trend has been enabled by an increase in available compute power; however, efforts to continue to scale electronic processors are impeded by the costs of communication, thermal management, power delivery and clocking. To improve scalability, we propose a digital optical neural network (DONN) with intralayer optical interconnects and reconfigurable input values. The path-length-independence of optical energy consumption enables information locality between a transmitter and a large number of arbitrarily arranged receivers, which allows greater flexibility in architecture design to circumvent scaling limitations. In a proof-of-concept experiment, we demonstrate optical multicast in the classification of 500 MNIST images with a 3-layer, fully-connected network. We also analyze the energy consumption of the DONN and find that digital optical data transfer is beneficial over electronics when the spacing of computational units is on the order of$$>10\,\upmu $$>10μm.

; ; ; ; ;
Award ID(s):
Publication Date:
Journal Name:
Scientific Reports
Nature Publishing Group
Sponsoring Org:
National Science Foundation
More Like this
  1. Abstract

    We present a proof of concept for a spectrally selective thermal mid-IR source based on nanopatterned graphene (NPG) with a typical mobility of CVD-grown graphene (up to 3000$$\hbox {cm}^2\,\hbox {V}^{-1}\,\hbox {s}^{-1}$$cm2V-1s-1), ensuring scalability to large areas. For that, we solve the electrostatic problem of a conducting hyperboloid with an elliptical wormhole in the presence of anin-planeelectric field. The localized surface plasmons (LSPs) on the NPG sheet, partially hybridized with graphene phonons and surface phonons of the neighboring materials, allow for the control and tuning of the thermal emission spectrum in the wavelength regime from$$\lambda =3$$λ=3to 12$$\upmu$$μm by adjusting the size of and distance between the circular holes in a hexagonal or square lattice structure. Most importantly, the LSPs along with an optical cavity increase the emittance of graphene from about 2.3% for pristine graphene to 80% for NPG, thereby outperforming state-of-the-art pristine graphene light sources operating in the near-infrared by at least a factor of 100. According to our COMSOL calculations, a maximum emission power per area of$$11\times 10^3$$11×103W/$$\hbox {m}^2$$m2at$$T=2000$$T=2000K for a bias voltage of$$V=23$$V=23V is achieved by controlling the temperature of the hot electrons through the Joule heating. By generalizing Planck’s theory to any grey body and derivingmore »the completely general nonlocal fluctuation-dissipation theorem with nonlocal response of surface plasmons in the random phase approximation, we show that the coherence length of the graphene plasmons and the thermally emitted photons can be as large as 13$$\upmu$$μm and 150$$\upmu$$μm, respectively, providing the opportunity to create phased arrays made of nanoantennas represented by the holes in NPG. The spatial phase variation of the coherence allows for beamsteering of the thermal emission in the range between$$12^\circ$$12and$$80^\circ$$80by tuning the Fermi energy between$$E_F=1.0$$EF=1.0eV and$$E_F=0.25$$EF=0.25eV through the gate voltage. Our analysis of the nonlocal hydrodynamic response leads to the conjecture that the diffusion length and viscosity in graphene are frequency-dependent. Using finite-difference time domain calculations, coupled mode theory, and RPA, we develop the model of a mid-IR light source based on NPG, which will pave the way to graphene-based optical mid-IR communication, mid-IR color displays, mid-IR spectroscopy, and virus detection.

    « less
  2. Abstract

    Emergent trends in the device development for neural prosthetics have focused on establishing stimulus localization, improving longevity through immune compatibility, reducing energy re-quirements, and embedding active control in the devices. Ultrasound stimulation can single-handedly address several of these challenges. Ultrasonic stimulus of neurons has been studied extensively from 100 kHz to 10 MHz, with high penetration but less localization. In this paper, a chip-scale device consisting of piezoelectric Aluminum Nitride ultrasonic transducers was engineered to deliver gigahertz (GHz) ultrasonic stimulus to the human neural cells. These devices provide a path towards complementary metal oxide semiconductor (CMOS) integration towards fully controllable neural devices. At GHz frequencies, ultrasonic wavelengths in water are a few microns and have an absorption depth of 10–20 µm. This confinement of energy can be used to control stimulation volume within a single neuron. This paper is the first proof-of-concept study to demonstrate that GHz ultrasound can stimulate neuronsin vitro. By utilizing optical calcium imaging, which records calcium ion flux indicating occurrence of an action potential, this paper demonstrates that an application of a nontoxic dosage of GHz ultrasonic waves$$(\ge 0.05\frac{W}{c{m}^{2}})$$(0.05Wcm2)caused an average normalized fluorescence intensity recordings >1.40 for the calcium transients. Electrical effects due to chip-scale ultrasound delivery wasmore »discounted as the sole mechanism in stimulation, with effects tested atα = 0.01 statistical significance amongst all intensities and con-trol groups. Ionic transients recorded optically were confirmed to be mediated by ion channels and experimental data suggests an insignificant thermal contributions to stimulation, with a predicted increase of 0.03oCfor$$1.2\frac{W}{c{m}^{2}}\cdot $$1.2Wcm2This paper paves the experimental framework to further explore chip-scale axon and neuron specific neural stimulation, with future applications in neural prosthetics, chip scale neural engineering, and extensions to different tissue and cell types.

    « less
  3. Abstract

    Thin film evaporation is a widely-used thermal management solution for micro/nano-devices with high energy densities. Local measurements of the evaporation rate at a liquid-vapor interface, however, are limited. We present a continuous profile of the evaporation heat transfer coefficient ($$h_{\text {evap}}$$hevap) in the submicron thin film region of a water meniscus obtained through local measurements interpreted by a machine learned surrogate of the physical system. Frequency domain thermoreflectance (FDTR), a non-contact laser-based method with micrometer lateral resolution, is used to induce and measure the meniscus evaporation. A neural network is then trained using finite element simulations to extract the$$h_{\text {evap}}$$hevapprofile from the FDTR data. For a substrate superheat of 20 K, the maximum$$h_{\text {evap}}$$hevapis$$1.0_{-0.3}^{+0.5}$$1.0-0.3+0.5 MW/$$\text {m}^2$$m2-K at a film thickness of$$15_{-3}^{+29}$$15-3+29 nm. This ultrahigh$$h_{\text {evap}}$$hevapvalue is two orders of magnitude larger than the heat transfer coefficient for single-phase forced convection or evaporation from a bulk liquid. Under the assumption of constant wall temperature, our profiles of$$h_{\text {evap}}$$hevapand meniscus thickness suggest that 62% of the heat transfer comes from the region lying 0.1–1 μm from the meniscus edge, whereas just 29% comes from the next 100 μm.

  4. Abstract

    Two-dimensional electron systems subjected to high transverse magnetic fields can exhibit Fractional Quantum Hall Effects (FQHE). In the GaAs/AlGaAs 2D electron system, a double degeneracy of Landau levels due to electron-spin, is removed by a small Zeeman spin splitting,$$g \mu _B B$$gμBB, comparable to the correlation energy. Then, a change of the Zeeman splitting relative to the correlation energy can lead to a re-ordering between spin polarized, partially polarized, and unpolarized many body ground states at a constant filling factor. We show here that tuning the spin energy can produce fractionally quantized Hall effect transitions that include both a change in$$\nu$$νfor the$$R_{xx}$$Rxxminimum, e.g., from$$\nu = 11/7$$ν=11/7to$$\nu = 8/5$$ν=8/5, and a corresponding change in the$$R_{xy}$$Rxy, e.g., from$$R_{xy}/R_{K} = (11/7)^{-1}$$Rxy/RK=(11/7)-1to$$R_{xy}/R_{K} = (8/5)^{-1}$$Rxy/RK=(8/5)-1, with increasing tilt angle. Further, we exhibit a striking size dependence in the tilt angle interval for the vanishing of the$$\nu = 4/3$$ν=4/3and$$\nu = 7/5$$ν=7/5resistance minima, including “avoided crossing” type lineshape characteristics, and observable shifts of$$R_{xy}$$Rxyat the$$R_{xx}$$Rxxminima- the latter occurring for$$\nu = 4/3, 7/5$$ν=4/3,7/5and the 10/7. The results demonstrate both size dependence and the possibility, not just of competition between different spin polarized states at the same$$\nu$$νand$$R_{xy}$$Rxy, but also the tilt- or Zeeman-energy-dependent- crossover between distinct FQHE associated withmore »different Hall resistances.

    « less
  5. Abstract

    Continuous multi-channel monitoring of biopotential signals is vital in understanding the body as a whole, facilitating accurate models and predictions in neural research. The current state of the art in wireless technologies for untethered biopotential recordings rely on radiative electromagnetic (EM) fields. In such transmissions, only a small fraction of this energy is received since the EM fields are widely radiated resulting in lossy inefficient systems. Using the body as a communication medium (similar to a ’wire’) allows for the containment of the energy within the body, yielding order(s) of magnitude lower energy than radiative EM communication. In this work, we introduce Animal Body Communication (ABC), which utilizes the concept of using the body as a medium into the domain of untethered animal biopotential recording. This work, for the first time, develops the theory and models for animal body communication circuitry and channel loss. Using this theoretical model, a sub-inch$$^3$$3[1″ × 1″ × 0.4″], custom-designed sensor node is built using off the shelf components which is capable of sensing and transmitting biopotential signals, through the body of the rat at significantly lower powers compared to traditional wireless transmissions. In-vivo experimental analysis proves that ABC successfully transmits acquired electrocardiogram (EKG) signals through the bodymore »with correlation$$>99\%$$>99%when compared to traditional wireless communication modalities, with a 50$$\times$$×reduction in power consumption.

    « less