Search

Streaming automatic speech recognition (ASR) is very important for many real-world ASR applications. However, a notable challenge for streaming ASR systems lies in balancing operational performance against latency constraint. Recently, a method of chunking, simulating future context and decoding, called CUSIDE, has been proposed for connectionist temporal classification (CTC) based streaming ASR, which obtains a good balance between reduced latency and high recognition accuracy. In this paper, we present CUSIDE-T, which successfully adapts the CUSIDE method over the recurrent neural network transducer (RNN-T) ASR architecture, instead of being based on the CTC architecture. We also incorporate language model rescoring in CUSIDE-T to further enhance accuracy, while only bringing a small additional latency. Extensive experiments are conducted over the AISHELL-1, WenetSpeech and SpeechIO datasets, comparing CUSIDE-T and U2++ (both based on RNN-T). U2++ is an existing counterpart of chunk based streaming ASR method. It is shown that CUSIDE-T achieves superior accuracy performance for streaming ASR, with equal settings of latency.

In weak-lensing cosmological studies, peak statistics is sensitive to nonlinear structures and thus complementary to cosmic shear two-point correlations. In this paper, we explore a new approach, namely, the peak steepness statistics, with the overall goal to understand the cosmological information embedded there in comparison with the commonly used peak height statistics. We perform the analyses with ray-tracing simulations considering different sets of cosmological parameters $\Omega_{\rm m}$ and $\sigma_8$. A theoretical model to calculate the abundance of high peaks based on steepness is also presented, which can well describe the main trend of the peak distribution from simulations. We employ $\Delta\chi^2$ and Fisher analyses to study the cosmological dependence of the two peak statistics using our limited sets of simulations as well as our theoretical model. Within our considerations without including potential systematic effects, the results show that the steepness statistics tends to have higher sensitivities to the cosmological parameters than the peak height statistics and this advantage is diluted with the increase of the shape noise. Using the theoretical model, we investigate the physical reasons accounting for the different cosmological information embedded in the two statistics. Our analyses indicate that the projection effect from large-scale structures plays an important role to enhance the gain from the steepness statistics. The redshift and cosmology dependence of dark matter halo density profiles also contributes to the differences between the two statistics., Comment: 20 pages, 15 figures. Accepted for publication in MNRAS

Remote sensing of the Earth's surface water is critical in a wide range of environmental studies, from evaluating the societal impacts of seasonal droughts and floods to the large-scale implications of climate change. Consequently, a large literature exists on the classification of water from satellite imagery. Yet, previous methods have been limited by 1) the spatial resolution of public satellite imagery, 2) classification schemes that operate at the pixel level, and 3) the need for multiple spectral bands. We advance the state-of-the-art by 1) using commercial imagery with panchromatic and multispectral resolutions of 30 cm and 1.2 m, respectively, 2) developing multiple fully convolutional neural networks (FCN) that can learn the morphological features of water bodies in addition to their spectral properties, and 3) FCN that can classify water even from panchromatic imagery. This study focuses on rivers in the Arctic, using images from the Quickbird, WorldView, and GeoEye satellites. Because no training data are available at such high resolutions, we construct those manually. First, we use the RGB, and NIR bands of the 8-band multispectral sensors. Those trained models all achieve excellent precision and recall over 90% on validation data, aided by on-the-fly preprocessing of the training data specific to satellite imagery. In a novel approach, we then use results from the multispectral model to generate training data for FCN that only require panchromatic imagery, of which considerably more is available. Despite the smaller feature space, these models still achieve a precision and recall of over 85%. We provide our open-source codes and trained model parameters to the remote sensing community, which paves the way to a wide range of environmental hydrology applications at vastly superior accuracies and 2 orders of magnitude higher spatial resolution than previously possible., Comment: 21 pages, 10 figures, 3 tables

In this article, we establish a characterization of the set $M(B^{0,b}_{p,\infty}(\mathbb{R}^n))$ of all pointwise multipliers of Besov spaces $B^{0,b}_{p,\infty}(\mathbb{R}^n)$ with only logarithmic smoothness $b\in\mathbb{R}$ in the special cases $p=1$ and $p=\infty$. As applications of these two characterizations, we clarify whether or not the three concrete examples, namely characteristic functions of open sets, continuous functions defined by differences, and the functions $e^{ik\cdot x}$ with $k\in\mathbb{Z}^n$ and $x\in\mathbb{R}^n$, are pointwise multipliers of $B^{0,b}_{1,\infty}(\mathbb{R}^n)$ and $B^{0,b}_{\infty,\infty}(\mathbb{R}^n)$, respectively; furthermore, we obtain the explicit estimates of $\|e^{ik \cdot x}\|_{M(B^{0,b}_{1,\infty}(\mathbb{R}^n))}$ and $\|e^{ik \cdot x}\|_{M(B^{0,b}_{\infty,\infty}(\mathbb{R}^n))}$. In the case that $p\in(1,\infty)$, we give some sufficient conditions and some necessary conditions of the pointwise multipliers of $B^{0,b}_{p,\infty}(\mathbb{R}^n)$ and a complete characterization of $M(B^{0,b}_{p,\infty}(\mathbb{R}^n))$ is still open. However, via a different method, we are still able to accurately calculate $\|e^{ik \cdot x}\|_{M(B^{0,b}_{p,\infty}(\mathbb{R}^n))}$, $k\in\mathbb{Z}^n$, in this situation. The novelty of this article is that most of the proofs are constructive and these constructions strongly depend on the logarithmic structure of Besov spaces under consideration., Comment: 64 pages; Submitted

Pre-training is prevalent in nowadays deep learning to improve the learned model's performance. However, in the literature on federated learning (FL), neural networks are mostly initialized with random weights. These attract our interest in conducting a systematic study to explore pre-training for FL. Across multiple visual recognition benchmarks, we found that pre-training can not only improve FL, but also close its accuracy gap to the counterpart centralized learning, especially in the challenging cases of non-IID clients' data. To make our findings applicable to situations where pre-trained models are not directly available, we explore pre-training with synthetic data or even with clients' data in a decentralized manner, and found that they can already improve FL notably. Interestingly, many of the techniques we explore are complementary to each other to further boost the performance, and we view this as a critical result toward scaling up deep FL for real-world applications. We conclude our paper with an attempt to understand the effect of pre-training on FL. We found that pre-training enables the learned global models under different clients' data conditions to converge to the same loss basin, and makes global aggregation in FL more stable. Nevertheless, pre-training seems to not alleviate local model drifting, a fundamental problem in FL under non-IID data., Comment: Accepted to ICLR 2023

We propose and study an efficient photonic CDE method using MZI-based coherent optical matrix multiplication. It improves the compensation performance by about 60% when the tap-length is limited, and only 50% taps of the theoretical value is needed for photonic CDE with 1-dB penalty., Comment: 4 pages,5 figures, Conference paper, submitted to European Conference on Optical Communication (ECOC)