arXiv subjects

Alexander Song

Publications and source records attributed to Alexander Song.

6 recordsLinked to original sources

arXiv2026

Multi-channel free-space optical convolutions with incoherent light

Free-space optical systems are promising candidates for high performance computing and have been particularly successful in the implementation of large-scale convolutions. Convolutions are the key operation in convolutional layers, which are used extensively in modern neural networks, especially in the context of image/video processing and generation. These optical accelerators have demonstrated remarkable performance in both processing rates and energy efficiency. Prior approaches have primarily demonstrated convolutions from a single input channel to one or more output channels. We extend these methods to perform true multi-channel convolutions, where multiple input channels are convolved with their own sets of convolutional kernels onto output channels. We simulate this approach using both ray-tracing and angular spectrum propagation and find the approach is highly-scalable. We then experimentally implement a proof-of-concept prototype to demonstrate multi-channel free-space optical convolutions.

physics.optics↗

arXiv2024

Low-power scalable multilayer optoelectronic neural networks enabled with incoherent light

Optical approaches have made great strides towards the goal of high-speed, energy-efficient computing necessary for modern deep learning and AI applications. Read-in and read-out of data, however, limit the overall performance of existing approaches. This study introduces a multilayer optoelectronic computing framework that alternates between optical and optoelectronic layers to implement matrix-vector multiplications and rectified linear functions, respectively. Our framework is designed for real-time, parallelized operations, leveraging 2D arrays of LEDs and photodetectors connected via independent analog electronics. We experimentally demonstrate this approach using a system with a three-layer network with two hidden layers and operate it to recognize images from the MNIST database with a recognition accuracy of 92% and classify classes from a nonlinear spiral data with 86% accuracy. By implementing multiple layers of a deep neural network simultaneously, our approach significantly reduces the number of read-ins and read-outs required and paves the way for scalable optical accelerators requiring ultra low energy.

cs.ET↗

arXiv2022

Engineering wavefronts with machine learned structured polarization

Optical approaches for wavefront shaping traditionally rely on phase modulation through holographic techniques. Shaping the phase determines a wave's diffraction and hence its intensity distribution in space. We instead show that shaping the polarization introduces a novel framework that permits the spatial modulation of polarization to control wavefront propagation and resulting amplitude distributions. We develop two distinct computational phase retrieval approaches for calculating the required polarization transformations and experimentally validate these. The first method extends the established Gerchberg-Saxton algorithm, while the second employs machine learning optimization to determine optimal polarization patterns. By implementing both amplitude and polarization control simultaneously using a single polarization mask, our approach significantly reduces system complexity compared to traditional methods. Our experimental results demonstrate the potential of polarization-based wavefront shaping as an efficient alternative to conventional techniques, paving the way for applications in optical manipulation and imaging.

physics.optics↗

arXiv2022

Data Processing of Functional Optical Microscopy for Neuroscience

Functional optical imaging in neuroscience is rapidly growing with the development of new optical systems and fluorescence indicators. To realize the potential of these massive spatiotemporal datasets for relating neuronal activity to behavior and stimuli and uncovering local circuits in the brain, accurate automated processing is increasingly essential. In this review, we cover recent computational developments in the full data processing pipeline of functional optical microscopy for neuroscience data and discuss ongoing and emerging challenges.

eess.IV↗

arXiv2019

Semi-Supervised First-Person Activity Recognition in Body-Worn Video

Body-worn cameras are now commonly used for logging daily life, sports, and law enforcement activities, creating a large volume of archived footage. This paper studies the problem of classifying frames of footage according to the activity of the camera-wearer with an emphasis on application to real-world police body-worn video. Real-world datasets pose a different set of challenges from existing egocentric vision datasets: the amount of footage of different activities is unbalanced, the data contains personally identifiable information, and in practice it is difficult to provide substantial training footage for a supervised approach. We address these challenges by extracting features based exclusively on motion information then segmenting the video footage using a semi-supervised classification algorithm. On publicly available datasets, our method achieves results comparable to, if not better than, supervised and/or deep learning methods using a fraction of the training data. It also shows promising results on real-world police body-worn video.

eess.IV↗

arXiv2017

Explicit block-structures for block-symmetric Fiedler-like pencils

In the last decade, there has been a continued effort to produce families of strong linearizations of a matrix polynomial $P(\lambda)$, regular and singular, with good properties. As a consequence of this research, families such as the family of Fiedler pencils, the family of generalized Fiedler pencils (GFP), the family of Fiedler pencils with repetition, and the family of generalized Fiedler pencils with repetition (GFPR) were constructed. In particular, one of the goals was to find in these families structured linearizations of structured matrix polynomials. For example, if a matrix polynomial $P(\lambda)$ is symmetric (Hermitian), it is convenient to use linearizations of $P(\lambda)$ that are also symmetric (Hermitian). Both the family of GFP and the family of GFPR contain block-symmetric linearizations of $P(\lambda)$, which are symmetric (Hermitian) when $P(\lambda)$ is. Now the objective is to determine which of those structured linearizations have the best numerical properties. The main obstacle for this study is the fact that these pencils are defined implicitly as products of so-called elementary matrices. In this paper we consider the family of block-minimal bases pencils, whose pencils are defined in terms of their block-structure, as a source of canonical forms for block-symmetric pencils. More precisely, we present four families of block-symmetric pencils which, under some generic nonsingularity conditions are block minimal bases pencils and strong linearizations of a matrix polynomial. We show that the block-symmetric GFP and GFPR, after some row and column permutations, belong to the union of these four families. Hence, these four families of pencils provide an alternative but explicit approach to the block-symmetric Fiedler-like pencils existing in the literature.

math.NA↗