José Figueiredo
Professor Auxiliar
Doctorate Program in Applied and Engineering Physics (DAEPHYS)
Programa de trabalho/Working programme
Título do programa de trabalhos/Title of the working programme
Neuromorphic nano-optoelectronic circuits for signal processing
Domínio Científico Principal/Main Scientific Domain
Applied and Engineering Physics
Título do programa de trabalhos/Title of the working programme
Neuromorphic nano-optoelectronic circuits for signal processing
Domínio Científico Principal/Main Scientific Domain
Applied and Engineering Physics
Domí¬nio Cientí¬fico/Scientific Domain
Optoelectronics
1. Sumário/Abstract
To realize the enormous potential of artificial intelligence (AI) and machine
learning information processing systems must be able to deal with astonishingly
large amounts of data in real time and allow ultra-wideband information exchange
within and between machines/equipment and systems. The growing demand for
low energy consuming computing systems capable of ultrafast data processing
and ultra-wideband communications requires solutions based on disruptive scientific
concepts and technologies beyond the ones the current semiconductor chipsets
and architectures are able to deliver.
Neuromorphic technologies capable to mimic neuro-biological architectures
present in the human nervous system and brain seems to be the most promising
approaches to fulfil the requirements of AI and of the future information
and communication systems, due to their capability for incorporate real-time
learning, plasticity, problem-solving abilities and enabling evolutionary
changes.
Purely neuromorphic semiconductor electronic approaches, where billions
of transistors need to be arranged in much more advanced than today’s configurations
to work in parallel in what is called inter-neuron connections (like synapses
in the human central nervous system and brain), seems to have intrinsic performance
bottlenecks in terms of operation speed and interconnectivity.
Optoelectronic and photonic neuromorphic technologies offer much more versatile
solutions with significant performance gains, particularly in terms of speed
and amount of data processing, and energy consumption, since huge amounts
of parallel streams of data can be processed and carried by light signals
at different but close wavelengths, using a small fraction of the energy needed
by electrical pulses in present semiconductor chipset.
This work program aims the investigation of novel neuromorphic nano-optoelectronic
circuits that combine resonant tunnelling nanoelectronic structures with photonic
devices. The main focus will be the development of brain-inspired (neuromorphic)
compact optoelectronic circuits capable of emulate some of the features of
spiking neurons and dynamic synapses. We believe that brain-inspired optoelectronic
circuits will display considerable superior performances in terms of speed,
operating bandwidth and energy consumption to the purely electronic neuromorphic
signal processing solutions.
2. Estado da Arte/ State of the art
The so called industry 4.0 (also commonly referred to as the fourth industrial
revolution), particularly the pillars associated to the internet of everything
concept (“the intelligent connection of people, process, data and things”)
will give rise to a myriad of gadget-like devices and equipment whose full
potential operation will be dependent on the exchange of great amounts of
data and internet traffic levels many folds above the ones required today.
These demands for high information processing speed, parallel processing
and ultra-wide bandwidth, will require novel technological solutions with
tight energy consuming requirements. Moreover, many of these equipment and
machines will incorporate great levels of artificial intelligence (AI). As
consequence there is the need of development of highly efficient ultrafast
information processing schemes and ultrahigh bandwidth communication solutions
within the circuits/equipment blocks and between equipment’s and systems.
Brain-like (neuromorphic) inspired information processing technologies,
including computing and communication, seem to have all the requirements
to comply with the demands of ultrahigh speed and bandwidth. However, while
the implementation of AI systems using computer algorithms of neural networks
is emerging rapidly, the development of the hardware basic elements of an
artificial brain, specifically neuromorphic microchips, is just taking the
first steps.
Relying on the tremendous success of the microelectronics, it would be expected
that semiconductor electronics would be easiest approach towards the implementation
of neuromorphic systems. However, in order to achieve this goal semiconductor
electronics needs to overcome some bottleneck the actual technologies are
facing because of the fundamental physics limitations associated to the electric
field control of charge flows within current semiconductor chipset implementations.
Light based brain-like (neuromorphic) information processing approach seems
to offer significant performance gains, since many parallel streams of data
can be processed and carried by light signals of different wavelengths using
a fraction of the energy than electrical pulses. Therefore, the scientific
and engineering communities foresee that light based brain-inspired solutions
that take advantage of lightwave characteristics (including electromagnetic
interference immunity and the enormous bandwidth available) will bring to
life disruptive technologies with considerable superior performance in terms
of operation speed, bandwidth and energy consumption [1,2]. Intense research
efforts are underway throughout the world towards the demonstration of light
based neuromorphic technologies capable of emulating the brain spike-encoded
basic information processing procedures [2,3].
3. Objectivos/Objectives
The aim of the doctoral project is to develop brain-inspired neuromorphic
optoelectronic circuits capable of emulate some of the features of spiking
neurons and dynamic synapses, which are based in the integration of resonant
tunnelling diode (RTD) structures with optoelectronic devices. The realization
of the working program will comprehend the study of novel neuromorphic nano-optoelectronic
circuits that combine nanoelectronic devices [e.g. semiconductor double barrier
quantum well resonant tunnelling diodes structures (DBQW-RTDs)] with photonic
devices [e.g. micro/nano light emitting diodes (LEDs) and/or micro/nano laser
diodes (LDs)].
DBQW-RTDs are nanoscale semiconductor electronic devices with a highly-nonlinear
current-voltage characteristics showing ultra-broadband (from dc up to 2 THz)
negative differential conductance (NDC), or electrical gain, at room temperature.
These properties allow their operation as excitable systems. Moreover, the
addition of light absorbing and confining layers to the DBQW-RTD structures
adds new functionalities to the device including optical modulation (RTD-OM)
and photo-detection (RTD-PD) which might lead to the implementation of very
fast electronic and optoelectronic circuits, such as optoelectronic oscillators,
whose operation can be controlled by both electrical and light signals.
Preliminary work on RTD/RTD-PD structures indicates that when integrated
with optoelectronic devices such as light sources (such as LEDs or LDs) they
allow the implementation of circuits with advanced properties including their
operation as optoelectronic excitable dynamic systems with interesting neuromorphic
functionalities. RTD/RTD-PD-LED/LD circuits can operate in several modes,
including (2): i) Bistable; ii) Self-sustained oscillations; iii) Excitable
dynamics; and v) Neural inhibition dynamics. Of particular importance is the
operation as an excitable system.
In the excitable regime appropriate perturbations (electrical and/or optical)
can trigger neuron-like excitable electrical spikes with around 1 V of amplitude
that can directly modulates LEDs/LDs, which generate neuron-like optical pulses.
The neuron-like electrical/optical spike pulses can then be used to drive
other identical neuromorphic circuits forming neural networks where the different
delays between circuits are used to control the functionalities of individual
devices/circuits such as resonators firing.
We believe that such brain-inspired compact neuromorphic optoelectronic
circuits will display considerable superior performances in terms of speed,
operating bandwidth and energy consumption when compared with neuromorphic
electronic signal processing solutions. The operation principle of the RTD-(PD)-LED/LD
neuromorphic photonic circuits under consideration takes advantage of the
RTD nonlinear non-monotonic current-voltage (I-V) characteristic, which shows
significant negative differential conductance (NDC) from dc up to terahertz
frequencies. For more details see [3].
4. Descrição detalhada/Detailed description
The work targets the investigation of neuromorphic nano-optoelectronic devices
and circuits based on the integration of RTDs with LEDs and/or LDs aiming
the development of brain-inspired circuits capable of emulate spiking neurons
and dynamic synapses. To achieve a level of knowledge that can lead to the
development and implementation of above described neuromorphic nano-optoelectronic
devices and circuits it is necessary to understand the operation principles
and the functionalities of their basic building blocks. This will require
knowledge (and benefits from the progresses) in multiple fields including
biology, physics, mathematics, photonics, computer science, electrical engineering,
and networks.
The work will start with the studies towards the familiarization of neuronal
properties and behaviours, particularly the ones associated to spiking neurons
and dynamic synapses. In parallel, the candidate will become familiar with
the principles of operation of RTDs/RTD-PDs and of prototype LEDs and LDs,
and with the relevant microelectronic integration techniques. At the same
time, the candidate will also acquire expertise to be able to use existing
numerical models and simulation packages of the devices currently implemented
using MATLAB. Particular attention will be given to the investigation of dynamics
of Fitzhugh-Nagumo excitable like systems with delayed coupling.
Description of the work plan tasks:
Task 1: State of the art and relevant physical and mathematical models
The activities will start with a thoroughly study of the concepts associated
to photonic and optoelectronic approaches towards the implementation of neuromorphic
information processing systems, including RTD implementation technologies
and RTD operation principle, light emitting diodes and laser diodes structures
and operation principles. Particular attention will be given to study of the
physical and mathematical models of the generation and transmission of a
stimulus through the human nervous system, including the FitzHugh–Nagumo and
Hodgkin–Huxley equations which models in a detailed manner activation and
deactivation dynamics of a spiking neuron. The work will include the improvement
of the exiting numerical models/simulation packages of the devices and the
circuits, and the development of new numerical models/simulation packages
capable to describe coupled resonators. MATLAB or other similar numerically
oriented programming language package will be used. Several RTD-LED epi-layer
structure designs will be investigated with the help of modelling and simulation
tools of advanced nanostructures and quantum transport simulators such as
Wingreen and nanoHub packages.
Task 2: Circuits implementation and characterization
The activities will also include detailed investigation of the operation
regimes of RTD circuits incorporating LEDs and LDs, both numerical/computational
and experimental, first using commercial LEDs and LDs, and then with prototype
designed to be integrated with RTDs. The work will then focus in the study
of circuits with time-delayed feedback mechanisms. The time-delayed feedback
implementation will consist of an optical delay line inserted off-chip, typically
using a low loss optical fibre, with a given time delay. This delay line will
provide a mechanism of reinjection of the fired optical pulses, analogous
to the autaptic neuron. This task will also comprehend the investigation of
the designs possibilities of micro/nano LEDs incorporating single (or multiple)
double barrier quantum well-like structures that show a light-current characteristic
that mimics the DBQW-RTD current-voltage characteristic non-linearities.
Task 3: Generation and propagation of neural spike
A thoroughly investigation of the firing dynamics of a single RTD-(PD)-LED/LD
resonator artificial neuron will be done. The investigation will comprehend
the generation of neural spike (when operating in the excitable regime) combined
with a mechanism of time-delayed feedback. RTD-PD-LED/LD operating in the
excitable dynamic regime combined with a mechanism of time-delayed feedback
will corresponds to the solid-state autaptic neuron configuration. The work
will include the study of the strategies that can allow the successful implantation
of circuit configurations that maximize efficiency and reduce the energy consumption,
therefore decreasing the amount of power dissipated on a single die, without
compromising their expected functionalities.
Task 4: Evaluation of interaction dynamics of artificial neurons
The task will comprehend the implementation and the detailed characterization
of the modes of operation of single and coupled nano-optoelectronic resonators
and the determination of operation conditions that allows the realization
of neuromorphic functions. The coupling between RTD-(PD)-LED/LD resonators
will be investigated in order to assess the practical feasibility of RTD-(PD)-LED/LD
based neural networks. The effects of competing phenomenon between gain and
loss when coupled RTD-(PD)-LED/LD resonators are operated in different conditions
will be evaluated. These studies will help us to understand how to manipulate
and fully tune the resonators excitation dynamics which are critical for achieving
high-level functions in neural networks.
Task 5: Thesis writing up
The final task will be the thesis writing up.
The above tasks do not necessarily follow a chronological order, with some
having activities that will be realized throughout all the period of the work
plan.
Interactions with other research centres/labs:
The project aims the investigation and development of novel neuromorphic
circuits capable of emulating certain brain/neural function based on the integration
of nanoelectronic devices with micro/nano photonic devices. The research
will be hosted by the CENTRA-SIM research labs at the Faculdade de Ciências
da Universidade de Lisboa. The work will benefit from and reinforce the interaction
of the supervisor research group with research groups or researcher from
the following institutions: Professor Edward Wasige group from the Electronic
and Nanoscale Engineering of School of Engineering, University of Glasgow;
Dr. Bruno Romeira, from the International Iberian Nanotechnology Laboratory,
Braga; Professor Paulo Monteiro from the Instituto de Telecomunicações,
Aveiro; and Dr. Tiago Alves from the Instituto de Telecomunicações,
Lisboa. It is also expected some degree of interaction with researchers of
the Grupo de Física Experimental, Instituto de Plasmas e Fusão
Nuclear (IPFN), Instituto Superior Tecnico, Universidade de Lisboa.
5. Referências/References
[1] “Deep learning with coherent nanophotonic circuits,” Yichen Shen, Nicholas
C. Harris, Scott Skirlo, Mihika Prabhu, Tom Baehr-Jones, Michael Hochberg,
Xin Sun, Shijie Zhao, Hugo Larochelle, Dirk Englund & Marin Soljačić,
Nat. Photonics 11, 441–446 (2017),
[2] “Neuromorphic Photonics,” P. R. Prucnal, B. J. Shastri, and M. C. Teich;
CRC Press/Taylor and Francis Group. ISBN-13: 978-1498725224
[3] “Delay dynamics of neuromorphic optoelectronic nanoscale resonators:
Perspectives and applications,” B. Romeira, J. M. L. Figueiredo, and J. Javaloyes,
Chaos: An Interdisciplinary Journal of Nonlinear Science: 27, 114323 (2017).
Supervisor
Name: José Figueiredo
Institution: FC/U Lisboa
Email: jose.figueiredo@ciencias.ulisboa.pt
ORCID ID: http://orcid.org/0000-0001-5668-7073
12 Março 2018