arXiv subjects

Karolis Misiunas

Publications and source records attributed to Karolis Misiunas.

7 recordsLinked to original sources

arXiv2023

Neural Architecture Search for Energy Efficient Always-on Audio Models

Mobile and edge computing devices for always-on classification tasks require energy-efficient neural network architectures. In this paper we present several changes to neural architecture searches (NAS) that improve the chance of success in practical situations. Our search simultaneously optimizes for network accuracy, energy efficiency and memory usage. We benchmark the performance of our search on real hardware, but since running thousands of tests with real hardware is difficult we use a random forest model to roughly predict the energy usage of a candidate network. We present a search strategy that uses both Bayesian and regularized evolutionary search with particle swarms, and employs early-stopping to reduce the computational burden. Our search, evaluated on a sound-event classification dataset based upon AudioSet, results in an order of magnitude less energy per inference and a much smaller memory footprint than our baseline MobileNetV1/V2 implementations while slightly improving task accuracy. We also demonstrate how combining a 2D spectrogram with a convolution with many filters causes a computational bottleneck for audio classification and that alternative approaches reduce the computational burden but sacrifice task accuracy.

eess.AS↗

arXiv2020

SEANet: A Multi-modal Speech Enhancement Network

We explore the possibility of leveraging accelerometer data to perform speech enhancement in very noisy conditions. Although it is possible to only partially reconstruct user's speech from the accelerometer, the latter provides a strong conditioning signal that is not influenced from noise sources in the environment. Based on this observation, we feed a multi-modal input to SEANet (Sound EnhAncement Network), a wave-to-wave fully convolutional model, which adopts a combination of feature losses and adversarial losses to reconstruct an enhanced version of user's speech. We trained our model with data collected by sensors mounted on an earbud and synthetically corrupted by adding different kinds of noise sources to the audio signal. Our experimental results demonstrate that it is possible to achieve very high quality results, even in the case of interfering speech at the same level of loudness. A sample of the output produced by our model is available at https://google-research.github.io/seanet/multimodal/speech.

eess.AS↗

arXiv2019

Nonlinear Electrophoresis of Highly Charged Nonpolarizable Particles

Nonlinear field dependence of electrophoresis in high fields has been investigated theoretically, yet experimental studies have failed to reach consensus on the effect. In this work, we present a systematic study on the nonlinear electrophoresis of highly charged submicron particles in applied electric fields of up to several kV/cm. First, the particles are characterized in the low-field regime at different salt concentrations and the surface charge density is estimated. Subsequently, we use microfluidic channels and video tracking to systematically characterize the nonlinear response over a range of field strengths. Using velocity measurements on the single particle level, we prove that nonlinear effects are present at electric fields and surface charge densities that are accessible in practical conditions. Finally, we show that nonlinear behavior leads to unexpected particle trapping in channels.

cond-mat.soft↗

arXiv2019

Density dependent speed-up of particle transport in channels

Collective transport through channels shows surprising properties under one-dimensional confinement: particles in a single file exhibit sub-diffusive behavior, while liquid confinement causes distance-independent correlations between the particles. Such interactions in channels are well studied for passive Brownian motion, but driven transport remains largely unexplored. Here, we demonstrate gating of transport due to a speed-up effect for actively driven particle transport through microfluidic channels. We prove that particle velocity increases with particle density in the channel due to hydrodynamic interactions under electrophoretic and gravitational forces. Numerical models demonstrate that the observed speed-up of transport originates from a hydrodynamic piston-like effect. Our discovery is fundamentally important for understanding protein channels and transport through porous materials and for designing novel sensors and filters.

cond-mat.soft↗

arXiv2018

QuipuNet: convolutional neural network for single-molecule nanopore sensing

Nanopore sensing is a versatile technique for the analysis of molecules on the single-molecule level. However, extracting information from data with established algorithms usually requires time-consuming checks by an experienced researcher due to inherent variability of solid-state nanopores. Here, we develop a convolutional neural network (CNN) for the fully automated extraction of information from the time-series signals obtained by nanopore sensors. In our demonstration, we use a previously published dataset on multiplexed single-molecule protein sensing. The neural network learns to classify translocation events with greater accuracy than previously possible, while also increasing the number of analysable events by a factor of five. Our results demonstrate that deep learning can achieve significant improvements in single molecule nanopore detection with potential applications in rapid diagnostics.

q-bio.QM↗

arXiv2015

Non-decaying hydrodynamic interactions along narrow channels

Particle-particle interactions are of paramount importance in every multi-body system as they determine the collective behaviour and coupling strength. Many well-known interactions like electro-static, van der Waals or screened Coulomb, decay exponentially or with negative powers of the particle spacing r. Similarly, hydrodynamic interactions between particles undergoing Brownian motion decay as 1/r in bulk, and are assumed to decay in small channels. Such interactions are ubiquitous in biological and technological systems. Here we confine two particles undergoing Brownian motion in narrow, microfluidic channels and study their coupling through hydrodynamic interactions. Our experiments show that the hydrodynamic particle-particle interactions are distance-independent in these channels. This finding is of fundamental importance for the interpretation of experiments where dense mixtures of particles or molecules diffuse through finite length, water-filled channels or pore networks.

physics.bio-ph↗

arXiv2014

Anisotropic diffusion of spherical particles in closely confining microchannels

We present here the measurement of the diffusivity of spherical particles closely confined by narrow microchannels. Our experiments yield a two-dimensional map of the position-dependent diffusion coefficients parallel and perpendicular to the channel axis with a resolution down to 129 nm. The diffusivity was measured simultaneously in the channel interior, the bulk reservoirs, as well as the channel entrance region. In the channel interior we found strongly anisotropic diffusion. While the perpendicular diffusion coefficient close to the confining walls decreased down to approximately 25% of the value on the channel axis, the parallel diffusion coefficient remained constant throughout the entire channel width. In addition to the experiment, we performed finite element simulations for the diffusivity in the channel interior and found good agreement with the measurements. Our results reveal the distinctive influence of strong confinement on Brownian motion, which is of significance to microfluidics as well as quantitative models of facilitated membrane transport.

physics.flu-dyn↗