THESIS
2021
1 online resource (xxiii, 130 pages) : illustrations (some color)
Abstract
Artificial neural networks (ANNs) have now been widely used for industrial applications
and also played more important roles in fundamental research. Although most
ANN hardware systems are electronically based, optical implementation is particularly
attractive because of its intrinsic parallelism and low energy consumption.
In this thesis, we demonstrate fully-functioned all-optical neural networks (AONNs),
in which the linear operations are programmed by spatial light modulators and Fourier
lenses, and the nonlinear optical activation functions are realized with electromagnetically
induced transparency (EIT) in laser-cooled atoms. We demonstrate a scalable AONN
with programmable linear operations and tunable nonlinear activation functions. Such
an AONN is scalable because all the erro...[
Read more ]
Artificial neural networks (ANNs) have now been widely used for industrial applications
and also played more important roles in fundamental research. Although most
ANN hardware systems are electronically based, optical implementation is particularly
attractive because of its intrinsic parallelism and low energy consumption.
In this thesis, we demonstrate fully-functioned all-optical neural networks (AONNs),
in which the linear operations are programmed by spatial light modulators and Fourier
lenses, and the nonlinear optical activation functions are realized with electromagnetically
induced transparency (EIT) in laser-cooled atoms. We demonstrate a scalable AONN
with programmable linear operations and tunable nonlinear activation functions. Such
an AONN is scalable because all the errors from different optical neurons are independent.
Although all-optical deep neural networks with a few neurons have been experimentally
demonstrated with acceptable errors recently, the feasibility of large scale AONNs remains
unknown because error might accumulate inevitably with increasing number of neurons
and connections. We verify its scalability by measuring and analyzing errors propagating
from a single neuron to the entire network.
Moreover, our hardware system is reconfigurable for different applications without
the need of modifying the physical structure. We confirm its capability and feasibility in
machine learning by multiple tasks. The AONN successfully classify the order and disorder
phases of a typical statistic Ising model. The demonstrated AONN scheme can be used to
construct various ANN architectures with the intrinsic optical parallel computation. The
feasibility of AONNs is further confirmed by recognizing handwritten digits and fashions
with classification rates of 81.8% and 71.3%, respectively.
The other application of AONNs is a regression task, quantum state tomography (QST). QST is a crucial ingredient for almost all aspects of experimental quantum information
processing. As an analog of the“imaging” technique in the quantum settings,
QST is born to be a data science problem, where machine learning techniques, noticeably
neural networks, have been applied extensively. Here, we build an integrated all-optical
machine for neural network QST, based on an AONN. The experimental results show
that AONN can predict the phase parameter of the quantum state accurately. Given
that optics is highly desired for quantum interconnections, our AONN-QST may contribute
to realization of all-optical quantum networks and inspire the ideas combining
optical neuromorphic computing with quantum information studies.
Post a Comment