arXiv subjects

Aashish Kumar

Publications and source records attributed to Aashish Kumar.

4 recordsLinked to original sources

arXiv2022

Conciseness: An Overlooked Language Task

We report on novel investigations into training models that make sentences concise. We define the task and show that it is different from related tasks such as summarization and simplification. For evaluation, we release two test sets, consisting of 2000 sentences each, that were annotated by two and five human annotators, respectively. We demonstrate that conciseness is a difficult task for which zero-shot setups with large neural language models often do not perform well. Given the limitations of these approaches, we propose a synthetic data generation method based on round-trip translations. Using this data to either train Transformers from scratch or fine-tune T5 models yields our strongest baselines that can be further improved by fine-tuning on an artificial conciseness dataset that we derived from multi-annotator machine translation test sets.

cs.CL↗

arXiv2021

Micrometer-thick, atomically random Si0.06Ge0.90Sn0.04 for silicon-integrated infrared optoelectronics

A true monolithic infrared photonics platform is within reach if strain and bandgap energy can be independently engineered in SiGeSn semiconductors. Herein, we investigate the structural and optoelectronic properties of a 1.5 {\mu}m-thick Si0.06Ge0.90Sn0.04 layer that is nearly lattice-matched to a Ge on Si substrate. Atomic-level studies demonstrate high crystalline quality and uniform composition and show no sign of short-range ordering and clusters. Room temperature spectroscopic ellipsometry and transmission measurements show direct bandgap absorption at 0.83 eV and a reduced indirect bandgap absorption at lower energies. Si0.06Ge0.90Sn0.04 photoconductive devices operating at room temperature exhibit dark current and spectral responsivity (1 A/W below 1.5 {\mu}m wavelengths) similar to Ge on Si devices, with the advantage of a near-infrared band gap tunable by alloy composition. These results underline the relevance of SiGeSn semiconductors in implementing a group IV material platform for silicon-integrated infrared optoelectronics.

cond-mat.mtrl-sci↗

arXiv2020

All-Group IV membrane room-temperature mid-infrared photodetector

Strain engineering has been a ubiquitous paradigm to tailor the electronic band structure and harness the associated new or enhanced fundamental properties in semiconductors. In this regard, semiconductor membranes emerged as a versatile class of nanoscale materials to control lattice strain and engineer complex heterostructures leading to the development of a variety of innovative applications. Herein we exploit this quasi-two-dimensional platform to tune simultaneously the lattice parameter and bandgap energy in group IV GeSn semiconductor alloys. As Sn content is increased to reach a direct band gap, these semiconductors become metastable and typically compressively strained. We show that the release and transfer of GeSn membranes lead to a significant relaxation thus extending the absorption wavelength range deeper in the mid-infrared. Fully released Ge$_{0.83}$Sn$_{0.17}$ membranes were integrated on silicon and used in the fabrication of broadband photodetectors operating at room temperature with a record wavelength cutoff of 4.6 $\mu$m, without compromising the performance at shorter wavelengths down to 2.3 $\mu$m. These membrane devices are characterized by two orders of magnitude reduction in dark current as compared to devices processed from as-grown strained epitaxial layers. The latter exhibit a content-dependent, shorter wavelength cutoff in the 2.6-3.5 $\mu$m range, thus highlighting the role of lattice strain relaxation in shaping the spectral response of membrane photodetectors. This ability to engineer all-group IV transferable mid-infrared photodetectors lays the groundwork to implement scalable and flexible sensing and imaging technologies exploiting these integrative, silicon-compatible strained-relaxed GeSn membranes.

physics.app-ph↗

arXiv2019

Attack Agnostic Statistical Method for Adversarial Detection

Deep Learning based AI systems have shown great promise in various domains such as vision, audio, autonomous systems (vehicles, drones), etc. Recent research on neural networks has shown the susceptibility of deep networks to adversarial attacks - a technique of adding small perturbations to the inputs which can fool a deep network into misclassifying them. Developing defenses against such adversarial attacks is an active research area, with some approaches proposing robust models that are immune to such adversaries, while other techniques attempt to detect such adversarial inputs. In this paper, we present a novel statistical approach for adversarial detection in image classification. Our approach is based on constructing a per-class feature distribution and detecting adversaries based on comparison of features of a test image with the feature distribution of its class. For this purpose, we make use of various statistical distances such as ED (Energy Distance), MMD (Maximum Mean Discrepancy) for adversarial detection, and analyze the performance of each metric. We experimentally show that our approach achieves good adversarial detection performance on MNIST and CIFAR-10 datasets irrespective of the attack method, sample size and the degree of adversarial perturbation.

cs.LG↗