PaperlayerView exact revision on arXiv

Source-linked AI summary

ITO-based Electro-absorption Modulator for Photonic Neural Activation Function

Rubab Amin, Jonathan George, Shuai Sun, Thomas Ferreira de Lima, Alexander N. Tait, Jacob Khurgin, Mario Miscuglio, Bhavin J. Shastri, Paul. R. Prucnal, Tarek El-Ghazawi, Volker J. Sorger

arXiv:1906.00017v1physics.app-phphysics.optics

TL;DR

Photonic circuits perform weighted sums efficiently, but neural-network activation calls for a suitable nonlinear response, and integrated all-optical nonlinear approaches remain challenging and inefficient. The paper demonstrates an ITO electro-absorption modulator driven by a photodiode as a photonic-neuron activation function. In the MNIST network, the modulator supports inference accuracy up to 97%.

  • Problem

    Photonic circuits can perform weighted sums efficiently, but activation requires nonlinear optics or an electro-optic module with an appropriate dynamic range.

    Source: §I · pp. 1–2

  • Method

    An ITO electro-absorption modulator uses a photodiode’s integrated weighted products and photovoltage to provide a photonic neuron’s nonlinear activation.

    Source: §I · p. 2

  • Results

    Up to 97% inference accuracy was reported for a two-layer, 100-node-per-layer MNIST network, with throughput of 5 GHz and power below 10 watts.

    Source: §IV · p. 8

  • Takeaways & Limitations

    The ITO modulator is presented as suitable for neural-network activation, with higher dynamic range and efficiency than similar silicon devices and a small power and footprint requirement.

    Source: §I · p. 2

Abstract

from arXiv · show

Recently integrated optics has become an intriguing platform for implementing machine learning algorithms and inparticular neural networks. Integrated photonic circuits can straightforwardly perform vector-matrix multiplicationswith high efficiency and low power consumption by using weighting mechanism through linear optics. Although,this can not be said for the activation function which requires either nonlinear optics or an electro-optic module withan appropriate dynamic range. Even though all-optical nonlinear optics is potentially faster, its current integrationis challenging and is rather inefficient. Here we demonstrate an electro-absorption modulator based on an IndiumTin Oxide layer, whose dynamic range is used as nonlinear activation function of a photonic neuron. The nonlinearactivation mechanism is based on a photodiode, which integrates the weighed products, and whose photovoltage drivesthe elecro-absorption modulator. The synapse and neuron circuit is then constructed to execute a 200-node MNISTclassification neural network used for benchmarking the nonlinear activation function and compared with an equivalentelectronic module.

I. INTRODUCTION

Photonic circuits efficiently perform weighted sums, but neural-network activation requires a suitable nonlinear response. The paper motivates ITO as an electro-absorption material and evaluates its modulator as an activation function for an MNIST network.

I. INTRODUCTION

  • Photonic circuits combine signals efficiently for weighted sums and vector-matrix operations, while wavelength-division multiplexing enables multiple wavelengths to share a physical channel.Photonic signals can also carry information through changes in phase, amplitude, or polarization.Source: §I · p. 1
  • Activation functions discriminate data and suppress noise, but electro-optic and photodiode-connected absorptive modulators trade baud rate against energy efficiency because of RC latency and electro-optic conversion.These modulators offer greater controllability, reconfigurability, reliability, and ease of integration than more exotic designs.Source: §I · p. 1
  • Modulator performance depends on both device configuration and active-material response; silicon’s low carrier tunability can require longer devices for the desired activation range.Plasmonic structures and photonic cavities can improve light-matter interaction, with plasmonic designs also incurring insertion losses.Source: §I · p. 2
  • ITO is presented as a candidate active material because it supports strong index modulation, ENZ behavior in the telecommunications band, and reliable integration processes.The passage also notes ITO deposition advances and monolithic hybrid integration.Source: §I · p. 2
  • The trained 200-node MNIST network reached 98% accuracy including noise using the fabricated ITO modulator’s non-ideal, asymmetric transfer function.The network has two layers of 100 nodes each.Source: §I · p. 2
  • ITO-based modulators are reported to provide greater dynamic range and efficiency than similar silicon devices, with a small power and footprint requirement for neural-network activation.The paper presents them as suitable for electro-optic neural networks.Source: §I · p. 2

II. METHODS

The methods examine free-carrier absorption in ITO and compare it with silicon carrier dynamics as a basis for modeling absorption.

II. METHODS

  • The study analyzes ITO electro-optical absorption dynamics driven by free carriers to establish absorption as a function of injected carriers and bias voltage.The analysis is intended to determine the absorption model used in the subsequent work.Source: §II · p. 2
  • ITO and silicon are compared because both use free-carrier absorption arising from carrier accumulation or depletion as their modulation mechanism.Source: §II · p. 2
  • The comparison uses the shared carrier mechanism to examine the absorption dynamics underlying modulation in both materials.Source: §II · p. 2

A. Free carrier absorption dynamics

The section models free-carrier absorption in waveguide modulators and compares ITO with silicon. It reports stronger absorption modulation per unit charge for ITO and discusses how carrier density and electrostatic gating shape the response.

A. Free carrier absorption dynamics

  • The model relates free-carrier absorption to planar carrier concentration and modal effective thickness, with absorption energy tied to the imaginary permittivity.The modal effective thickness measures optical-mode confinement to the active layer; the Lorentzian line shape reaches its maximum at resonance.Source: A. Free carrier absorption dynamics · pp. 2–3
  • ITO provides more modulation per unit charge than high-quality silicon, and its absorption-versus-density curve is several orders per decade steeper.The section attributes silicon’s limited attainable absorption to its low broadening parameter, even at higher carrier concentrations.Source: A. Free carrier absorption dynamics · p. 4
  • The comparison uses silicon parameters of 0.26m0 effective mass and 1100 cm2V −1s−1 mobility, and ITO parameters including γ = 1.8×1014rads−1 and 0.35m0 effective mass.The passage notes that ITO properties depend on deposition conditions, defect states, and film thickness.Source: A. Free carrier absorption dynamics · p. 4
  • ITO’s epsilon-near-zero condition is described as reachable at telecom-relevant 1550 nm, whereas silicon’s stated ENZ condition is at 10 µm in the infrared.The text associates operation near ENZ with stronger modulation effects as optical interaction intensifies.Source: A. Free carrier absorption dynamics · p. 5

B. Neural Network

The network uses microring resonator banks for optical weighting and models photodiode shot and thermal noise.

B. Neural Network

  • Microring resonator banks provide optical weights bounded between -1 and 1 in a push-pull configuration.The cited approach can potentially support more than 100 channels.Source: B. Neural Network · p. 5
  • The model includes additive Gaussian noise and photodiode shot noise based on signal and dark currents.The dark current is taken as 0.05 nA.Source: B. Neural Network · p. 5
  • Thermal noise is modeled with room temperature set to 300 K and equivalent resistance set to 50 Ω.Source: B. Neural Network · p. 5

1. Training

The Keras models are trained with Adagrad and categorical cross-entropy, while training noise is reduced relative to inference noise.

1. Training

  • Training uses Keras with Adagrad, categorical cross-entropy, a learning rate of 0.005, and zero decay.Source: §1 · p. 6
  • Models are trained for 500 epochs with a batch size of 1024.Source: §1 · p. 6
  • Noise power during training is reduced to 10% of final inference noise to aid convergence while avoiding overfitting to a noiseless activation.Source: §1 · p. 6

2. Parameter Optimization

The photonic network has analog noise and depth-dependent bias and dynamic-range shifts; an optimization procedure centers the activation range at each layer.

2. Parameter Optimization

  • Unlike the digital network, the photonic network has analog shot and thermal noise, no quantization noise, and cascading noise.Source: §2 · p. 6
  • When activation saturates, the analog network can regenerate signals and act as a quasi-digital system in some configurations.Source: §2 · p. 6
  • To counter depth-driven bias and range shifts, the procedure finds the steepest activation point, expands to where slope falls below 0.05, then applies a DC bias to center the range.The activation is swept across 200 input-voltage points without noise.Source: §2 · p. 6

A. ITO electro-absorption activation function: Inference results

In simulated MNIST inference, the ITO activation exceeds 90% accuracy above 5 mW, while the same-footprint Si modulator remains below 30% even at 15 mW.

A. ITO electro-absorption activation function: Inference results

B. Experimental characterization of an ITO based Electro-Absorption Modulator Device

The ITO electro-absorption modulator is fabricated on a silicon waveguide and characterized at two device lengths. In the MNIST classifier, its experimentally fitted activation reaches over 97% accuracy at 15 mW, with the 5 µm device performing better below 5 mW.

B. Experimental characterization of an ITO based Electro-Absorption Modulator Device

  • The devices use a 10 nm ITO layer above a silicon waveguide, separated by SiO2 and capped with a gold contact.The waveguide measures 800 nm × 340 nm; the two modulator lengths are 5 and 20 µm.Source: B. Experimental characterization of an… · p. 6, Figure 4 · p. 7
  • The 5 and 20 µm devices achieve extinction ratios of 5 dB and more than 20 dB, respectively; the 5 µm device has approximately 1 dB insertion loss.The ITO film is electrostatically gated to tune carrier density and the portion of the optical field absorbed by the film.Source: B. Experimental characterization of an… · pp. 6–7
  • Both modulator sizes achieve greater than 90% simulated MNIST accuracy at laser powers of 5 mW or higher.The plotted activations are fitted with a 10-degree clipped polynomial over a 1–15 mW laser-power sweep.Source: Figure 5 · p. 7
  • The ITO modulator is reported to have a striking modulation range for its footprint, a broadband response, and no need for cavity feedback.The paper compares it with state-of-the-art silicon and LiNbO3 traveling-wave modulators, but the supplied passage gives no numerical comparison values.Source: B. Experimental characterization of an… · p. 7
  • At 15 mW, both ITO modulator sizes asymptotically converge to accuracy greater than 97%, while the 5 µm device performs better below 5 mW.The activation functions are fitted to experimental voltage–absorption data for the MNIST classifier.Source: B. Experimental characterization of an… · p. 7, Figure 5 · p. 7

IV. CONCLUSIONS

The paper incorporates an ITO free-carrier absorption model as a neural-network activation function. The modeled network reaches up to 97% handwritten-classification accuracy, and experimental modulator data show higher performance than the model predicts.

IV. CONCLUSIONS

  • The ITO-based activation model reaches up to 97% accuracy on handwritten classification in a two-layer network with 100 nodes per layer.The reported throughput is 5 GHz, with latency described as low and a power budget below 10 watts.Source: §IV · p. 8
  • The modeled optical platform estimates 60 fJ/bit per electro-absorption module and 12 pJ/bit for 200 nonlinear activation modules, excluding I/O and laser power.The per-module inference consumption is estimated as approximately 60 fJ/bit using 1/2CV^2.Source: §IV · p. 8
  • Experimental data from a fabricated ITO modulator show higher performance than the model predicts, which does not account for device asymmetries and material nonidealities.The work describes free-carrier absorption dynamics in ITO and incorporates its absorption model as the network activation function.Source: §IV · pp. 7–8
Loading 1906.00017v1…