PaperlayerView exact revision on arXiv

Source-linked AI summary

All-Optical Machine Learning Using Diffractive Deep Neural Networks

Xing Lin, Yair Rivenson, Nezih T. Yardimci, Muhammed Veli, Mona Jarrahi, Aydogan Ozcan

arXiv:1804.08711v2cs.NEcs.LGphysics.comp-phphysics.optics

TL;DR

The scope of optical neural-network implementations remains a research question. This paper introduces trainable diffractive networks and experimentally demonstrates 3D-printed systems for handwritten-digit classification and imaging-lens functions at terahertz frequencies.

  • Problem

    The paper examines whether physics-inspired neural-network designs can be implemented optically, a question motivated by optical parallel computing.

    Source: Main Text · p. 3, Discussion · p. 13

  • Method

    The authors train passive diffractive layers to work collectively as an all-optical neural network.

    Source: Main Text · p. 2

  • Results

    Experiments with 3D-printed D2NNs demonstrated handwritten-digit classification and imaging-lens functions at terahertz frequencies.

    Source: Discussion · p. 10

  • Takeaways & Limitations

    The authors identify potential applications in all-optical image analysis, feature detection, object classification, and microscope or camera designs.

    Source: Discussion · p. 11

    The reported implementation has discrepancies associated with printing errors and alignment issues, as well as residual amplitude effects from absorption losses.

    Source: Results · pp. 7–8

Abstract

from arXiv · show

We introduce an all-optical Diffractive Deep Neural Network (D2NN) architecture that can learn to implement various functions after deep learning-based design of passive diffractive layers that work collectively. We experimentally demonstrated the success of this framework by creating 3D-printed D2NNs that learned to implement handwritten digit classification and the function of an imaging lens at terahertz spectrum. With the existing plethora of 3D-printing and other lithographic fabrication methods as well as spatial-light-modulators, this all-optical deep learning framework can perform, at the speed of light, various complex functions that computer-based neural networks can implement, and will find applications in all-optical image analysis, feature detection and object classification, also enabling new camera designs and optical components that can learn to perform unique tasks using D2NNs.

Main Text

The paper introduces the Diffractive Deep Neural Network (D2NN), an all-optical framework in which trained diffractive layers work together to perform learned functions. Simulations and experiments demonstrate its use for handwritten-digit classification and imaging-lens implementation.

Main Text

  • D2NN layers act as artificial neurons whose learned transmission or reflection coefficients modulate diffracted waves, and are adjusted by computer-based error back-propagation.The coefficients determine each point’s amplitude and phase contribution to the wave reaching the next layer.Source: Main Text · p. 2
  • After training, fixed D2NN designs can be fabricated as passive optical layers that perform their trained task through diffraction at the speed of light.The paper describes 3D printing and lithography as fabrication options.Source: Main Text · pp. 2–3
  • The D2NN design formed a fully connected network and achieved 91.75% classification accuracy on MNIST; a handwritten-digit classifier was experimentally demonstrated using 3D-printed transmission layers.Source: Main Text · p. 3
  • Five stacked transmissive layers with 0.45 million neurons and 4 mm axial spacing experimentally implemented an imaging lens, with fewer than 0.1 billion connections.The compact layer spacing is associated in the text with the lower connection count than in the digit-classifier D2NN.Source: Main Text · p. 3
  • The framework is presented as a scalable, power-efficient optical engine for image analysis, feature detection, object classification, and optical components with learned imaging functions.The paper suggests such components might enable new camera or microscope designs.Source: Main Text · pp. 3–4
  • Transmission- or reflection-based spatial light modulators could enable reconfigurable D2NNs for arbitrary functions, transfer learning, and adjustments based on new data or user feedback.Source: Main Text · p. 4

Results

The results assess trained D2NNs for handwritten-digit classification and unit-magnification imaging, reporting numerical and experimental performance. They also describe experimental resolution and performance limitations associated with fabrication, alignment, and absorption losses.

D2NN Architecture

  • Each phase-modulating layer generates secondary waves that diffract and interfere between layers, forming the complex wave that feeds the next layer.The network’s neuron phase values are iteratively trained against a target output using error back-propagation.Source: Results · pp. 4–5

D2NN trained for handwritten digit classification

  • 91.75% classification accuracy was achieved on 10,000 held-out MNIST test images, with output energy focused into the detector region for each target digit.The test images were not used for training or validation.Source: Results · p. 6
  • For 50 3D-printed handwritten digits, average output-plane intensity showed focused input-beam energy reaching a maximum signal at each digit’s assigned detector, demonstrating learning capability despite small discrepancies from numerical results.Source: Results · p. 7
  • In numerical tests, the digit classifier focused about 38–53% of output-plane energy onto the correct detector, versus about 34–40% experimentally, except digit 1 at about 60%.Printing errors, alignment issues, and difficulty reproducing void regions contributed to discrepancies between simulations and experiments.Source: Results · p. 7

D2NN trained as an imaging lens – a physical auto-encoder

  • The 3D-printed five-layer D2NN projected unit-magnification images of input patterns at 0.4 THz, demonstrating imaging-lens functionality as a physical auto-encoder.The network was trained on about 1,500 randomly selected ImageNet images.Source: Results · p. 8

Results

  • For 1 mm, 2 mm, and 3 mm pinholes, the output images had FWHM values of 1.5 mm, 1.4 mm, and 2.5 mm, respectively.These pinhole measurements evaluated the D2NN point spread function (Fig. 4B).Source: Results · p. 8
  • At 0.4 THz (0.75 mm wavelength in air), the experimental network resolved 1.8 mm line-widths, versus ~1.2 mm in numerical testing.The experimental resolution was slightly worse than during numerical testing.Source: Results · p. 8
  • 3D-printing errors, potential layer misalignments, and absorption-related residual amplitude modulation were identified as sources of experimental degradation, partly explaining the handwritten-digit classifier's degradation.The amplitude modulation partly deviated from the phase-only network design.Source: Results · pp. 8–9
  • Although the system's theoretical numerical aperture approaches 1 in air, the trained network used only part of its spatial-frequency bandwidth, attributed to the large-scale image features in its training set.The training images were randomly selected from ImageNet.Source: Results · p. 9
  • Higher-resolution imaging could be pursued by training with finer spatial features or changing the loss function to enhance spatial resolution.The proposed training choices could help the D2NN approach the system's theoretical diffraction-limited numerical aperture.Source: Results · p. 9

Discussion

The discussion examines D2NN inference as all-optical computation through passive diffraction, including practical constraints, ways to reduce losses and errors, possible extensions, and applications in imaging and object recognition. It also leaves open whether wave-based communication can outperform computer-based deep networks and considers virtual wave propagation as a possible means of altering diffraction laws to improve performance.

Discussion

  • At 0.4 THz, approximately 1-mm 3D-printed layers attenuated ~51% of power; thinner substrates, lower-loss materials, and antireflection coatings could reduce optical losses.With low-loss materials and suitable wavelengths, phase-only D2NN energy efficiency would be limited by Fresnel reflections at layer surfaces.Source: Discussion · pp. 9–10
  • D2NNs can extend to amplitude-only or mixed phase/amplitude transmissive or reflective designs, while added optical gain could boost diffracted signals at the cost of complexity.Amplifying layers could help manage the photon budget when placed after passive layers.Source: Discussion · pp. 10–11
  • Large-area fabrication and wide-field optics could scale D2NNs to tens or hundreds of millions of neurons and hundreds of billions of connections, enabling parallel, power-efficient processing.Soft-lithography is one example of a large-area fabrication method.Source: Discussion · p. 11
  • D2NNs integrated with lensfree on-chip imaging or detection could enable compact, field-portable systems and support all-optical image analysis, feature detection, object classification, and learned imaging tasks.A neural-network detection field of view >18 cm2 can be routinely achieved using state-of-the-art CCDs.Source: Discussion · p. 11
  • The holder’s estimated error from ideal neuron positions in a layer was below 0.1 mm, with minor network-performance effect; increasing layer area and footprint partly eases alignment requirements but increases system size and input optical-power needs.Source: Discussion · pp. 11–12
  • Combined diffractive-surface reconstruction errors, absorption-related losses at different layers, and 0.1 mm random misalignment for each network layer reduced digit-classification accuracy from 91.75% (Fig. 3C).The comparison contrasts Fig. 3C, a design without alignment errors or imperfections, with Fig. S4.Source: Discussion · p. 11
  • Spatial light modulators (SLMs) could make D2NNs reconfigurable for different tasks and help mitigate alignment errors or other optical imperfections, though they add system complexity.The discussion describes fully SLM-based and hybrid systems combining SLMs with fabricated passive layers.Source: Discussion · p. 12
  • The authors identify computer-based neural networks using D2NN mathematics as a future avenue, asking whether wave-propagation-inspired, coherent or partially coherent interlayer communication could offer machine-learning insights or capabilities.A virtual implementation could loosen the physical constraints on wave-function forms imposed by a 3D-printed D2NN.Source: Discussion · p. 13
  • Whether D2NNs’ wave-based communication between layers provides advantages over traditional computer-based deep networks remains to be rigorously explored.The discussion also suggests that virtual wave propagation in a computer can potentially alter optical diffraction laws to benefit network performance.Source: Discussion · p. 13

FIGURES AND CAPTIONS

The figures show D2NNs trained to classify handwritten digits and act as imaging lenses, then fabricated and experimentally tested at terahertz frequencies. They also depict the layered optical architecture and its coherent-wave computation.

D2NN architecture

  • Each D2NN layer uses trainable complex-valued transmission or reflection coefficients, and the fixed fabricated network performs its learned function at the speed of light.D2NN computation uses coherent complex-valued inputs, multiplicative bias terms, and free-space diffraction to determine interference between modulated secondary waves.Source: Figure 1 · p. 17

Experimental testing

  • The classifier and lens were implemented as five-layer 3D-printed D2NNs and tested in a THz setup using continuous-wave radiation at 0.4 THz.The setup caption also identifies the detection chain at the network output plane.Source: Figure 2 · p. 18, FIGURES AND CAPTIONS · p. 18

Handwritten digit classifier

  • 91.75% classification accuracy was achieved in numerical testing of 10,000 handwritten digits.The figure also shows experimental results for 50 printed digits, five per class.Source: FIGURES AND CAPTIONS · p. 19, Figure 3 · p. 19

Imaging lens

  • The imaging lens D2NN produced outputs for letter and pinhole inputs, resolved a 1.8 mm line width, and tolerated axial defocusing up to ~12 mm.Outputs for the letter and pinhole inputs are compared with free-space diffraction; similar output images were obtained at four input-plane locations for the same 3-mm pinhole.Source: Figure 4 · p. 20

Related papers

Loading 1804.08711v2…