scispace - formally typeset

Author

Pratyush Kumar

Bio: Pratyush Kumar is an academic researcher from Indian Institute of Technology Madras. The author has contributed to research in topic(s): Scheduling (computing) & Cyber-physical system. The author has an hindex of 20, co-authored 110 publication(s) receiving 1389 citation(s). Previous affiliations of Pratyush Kumar include École Polytechnique Fédérale de Lausanne & University of California, Santa Barbara.
Papers
More filters

Posted Content
Abstract: Recent methods in speech and language technology pretrain very LARGE models which are fine-tuned for specific tasks. However, the benefits of such LARGE models are often limited to a few resource rich languages of the world. In this work, we make multiple contributions towards building ASR systems for low resource languages from the Indian subcontinent. First, we curate 17,000 hours of raw speech data for 40 Indian languages from a wide variety of domains including education, news, technology, and finance. Second, using this raw speech data we pretrain several variants of wav2vec style models for 40 Indian languages. Third, we analyze the pretrained models to find key features: codebook vectors of similar sounding phonemes are shared across languages, representations across layers are discriminative of the language family, and attention heads often pay attention within small local windows. Fourth, we fine-tune this model for downstream ASR for 9 languages and obtain state-of-the-art results on 3 public datasets, including on very low-resource languages such as Sinhala and Nepali. Our work establishes that multilingual pretraining is an effective strategy for building ASR systems for the linguistically diverse speakers of the Indian subcontinent.

Journal ArticleDOI
Pratyush Kumar1, Abhishek Thakur1, Sandip K. Saha1, Atul Sharma1  +2 moreInstitutions (2)
Abstract: The lead-lithium cooled ceramic breeder uses spherical lithium titanate pebbles, which act as the breeder for generating tritium as the fuel for nuclear fusion reaction and helium as the purging gas. In this work, the flow-field of helium gas and temperature distribution inside a rectangular canister is investigated numerically, considering the pebble bed as the porous medium. In the numerical model, the local thermal non-equilibrium (LTNE) is assumed to exist between the pebbles and the purging gas, and a local porosity distribution is considered. Considering the radiation effect, a heterogeneous and LTNE based solver developed in OpenFOAM software is validated with the published results. Further, the effects of different heat generation rates, mass flow rates, and packing arrangements (random and uniform arrangements, i.e. BCC and FCC) on the flow-field and temperature profile are studied. The BCC and FCC packing arrangements have better purging behaviour than random packing. Further, the effect of inlet and outlet configuration is evaluated to find the optimal locations for uniform distributions of purging gas in the canister. With a higher value of area-weighted average velocity, the diagonally opposite inlet-outlet is the best choice for removing tritium from the packed bed.

Proceedings ArticleDOI
21 Oct 2021
Abstract: Digital hardware is verified by comparing its behavior against a reference model on a range of randomly generated input signals. The random generation of the inputs hopes to achieve sufficient coverage of the different parts of the design. However, such coverage is often difficult to achieve, amounting to large verification efforts and delays. An alternative is to use Reinforcement Learning (RL) to generate the inputs by learning to prioritize those inputs which can more efficiently explore the design under test. In this work, we present VeRLPy [3], an open-source library to allow RL-driven verification with limited additional engineering overhead. This contributes to two broad movements within the EDA community of (a) moving to open-source tool chains and (b) reducing barriers for development with Python support. We also demonstrate the use of VeRLPy for a few designs and establish its value over randomly generated input signals.

Posted Content
Abstract: AI technologies for Natural Languages have made tremendous progress recently. However, commensurate progress has not been made on Sign Languages, in particular, in recognizing signs as individual words or as complete sentences. We introduce OpenHands, a library where we take four key ideas from the NLP community for low-resource languages and apply them to sign languages for word-level recognition. First, we propose using pose extracted through pretrained models as the standard modality of data to reduce training time and enable efficient inference, and we release standardized pose datasets for 6 different sign languages - American, Argentinian, Chinese, Greek, Indian, and Turkish. Second, we train and release checkpoints of 4 pose-based isolated sign language recognition models across all 6 languages, providing baselines and ready checkpoints for deployment. Third, to address the lack of labelled data, we propose self-supervised pretraining on unlabelled data. We curate and release the largest pose-based pretraining dataset on Indian Sign Language (Indian-SL). Fourth, we compare different pretraining strategies and for the first time establish that pretraining is effective for sign language recognition by demonstrating (a) improved fine-tuning performance especially in low-resource settings, and (b) high crosslingual transfer from Indian-SL to few other sign languages. We open-source all models and datasets in OpenHands with a hope that it makes research in sign languages more accessible, available here at https://github.com/AI4Bharat/OpenHands .

Posted Content
Abstract: Task-agnostic pre-training followed by task-specific fine-tuning is a default approach to train NLU models. Such models need to be deployed on devices across the cloud and the edge with varying resource and accuracy constraints. For a given task, repeating pre-training and fine-tuning across tens of devices is prohibitively expensive. We propose SuperShaper, a task agnostic pre-training approach which simultaneously pre-trains a large number of Transformer models by varying shapes, i.e., by varying the hidden dimensions across layers. This is enabled by a backbone network with linear bottleneck matrices around each Transformer layer which are sliced to generate differently shaped sub-networks. In spite of its simple design space and efficient implementation, SuperShaper discovers networks that effectively trade-off accuracy and model size: Discovered networks are more accurate than a range of hand-crafted and automatically searched networks on GLUE benchmarks. Further, we find two critical advantages of shape as a design variable for Neural Architecture Search (NAS): (a) heuristics of good shapes can be derived and networks found with these heuristics match and even improve on carefully searched networks across a range of parameter counts, and (b) the latency of networks across multiple CPUs and GPUs are insensitive to the shape and thus enable device-agnostic search. In summary, SuperShaper radically simplifies NAS for language models and discovers networks that generalize across tasks, parameter constraints, and devices.

Cited by
More filters

Journal ArticleDOI
TL;DR: Following a taxonomy development approach, 22 empirically and conceptually grounded design dimensions contingent on chatbots’ temporal profiles are compiled, and three time-dependent chatbot design archetypes are abstracted: Ad-hoc Supporters, Temporary Assistants, and Persistent Companions.
Abstract: Users interact with chatbots for various purposes and motivations – and for different periods of time. However, since chatbots are considered social actors and given that time is an essential component of social interactions, the question arises as to how chatbots need to be designed depending on whether they aim to help individuals achieve short-, medium- or long-term goals. Following a taxonomy development approach, we compile 22 empirically and conceptually grounded design dimensions contingent on chatbots’ temporal profiles. Based upon the classification and analysis of 120 chatbots therein, we abstract three time-dependent chatbot design archetypes: Ad-hoc Supporters, Temporary Assistants, and Persistent Companions. While the taxonomy serves as a blueprint for chatbot researchers and designers developing and evaluating chatbots in general, our archetypes also offer practitioners and academics alike a shared understanding and naming convention to study and design chatbots with different temporal profiles.

Journal ArticleDOI
Yi-wen Zhang1Institutions (1)
TL;DR: A novel algorithm called EAU is presented, which applies the actual execution time to re-compute the utilization of the task when a job is completed early or is released and can save up to 46.84% of energy compared with existing algorithms.
Abstract: Many real-time applications with different criticalities are integrated into a single mixed-criticality (MC) real-time system. Previous studies on MC systems assume that all tasks are executed in their worst case execution time and ignore the energy consumption in the high-criticality mode. In this paper, we focus on the actual execution time of tasks and consider the energy consumption in both the low-criticality mode and the high-criticality mode. First, we present a novel algorithm called EAU, which applies the actual execution time to re-compute the utilization of the task when a job is completed early or is released. In addition, it can apply the slack time generated from the early completion jobs and the jobs for which the processor speed is lower than the maximum processor speed in the high-criticality mode. Secondly, we analyze the scheduling feasibility of EAU. Finally, experiments are conducted to evaluate the performance of the proposed algorithm. The experimental results show that EAU can save up to 46.84% of energy compared with existing algorithms.

1 citations



DOI
01 Jan 2022
Abstract: The task of hope speech detection has gained traction in the natural language processing field owing to the need for an increase in positive reinforcement online during the COVID-19 pandemic. Hope speech detection focuses on identifying texts among social media comments that could invoke positive emotions in people. Students and working adults alike posit that they experience a lot of work-induced stress further proving that there exists a need for external inspiration which in this current scenario, is mostly found online. In this paper, we propose a multilingual model, with main emphasis on Dravidian languages, to automatically detect hope speech. We have employed a stacked encoder architecture which makes use of language agnostic cross-lingual word embeddings as the dataset consists of code-mixed YouTube comments. Additionally, we have carried out an empirical analysis and tested our architecture against various traditional, transformer, and transfer learning methods. Furthermore a k-fold paired t test was conducted which corroborates that our model outperforms the other approaches. Our methodology achieved an F1-score of 0.61 and 0.85 for Tamil and Malayalam, respectively. Our methodology is quite competitive to the state-of-the-art methods. The code for our work can be found in our GitHub repository (https://github.com/arunimasundar/Hope-Speech-LT-EDI).

Book ChapterDOI
Vadim Kramar1, Juha Röning2, Juha Erkkila3, Henry Hinkula1  +2 moreInstitutions (3)
01 Jan 2022
Abstract: The European Union (EU) regulations regarding the unmanned aircraft system (UAS) that came into force in 2021 emphasise technological and operational safety. Those regulations have been developed on the common rules in the field of civil aviation and establishing a European Union Aviation Safety Agency (EASA). The implementation of the regulations and compliant UAS operator activities are still the ground of the future. Therefore, it is essential to systematically gather information about all the factors affecting UAS operations in a safe and meaningful manner. This book chapter introduces the Nordic as well as generic challenges for UAS operations. The challenges can be divided into two main categories: technological and operational. Based on the extensive literature review and authors’ practical experience, both types of challenges are grouped by relevance topics. For example, the weather-related phenomena challenges overlap in both technological and operational categories but still can be mitigated differently. Technological challenges are usually mitigated by UAS design and human-computer interactions, while operational challenges may be mitigated with legislation and organisational activities and personal UAS operator qualities. Finally, the needs for further research on the challenges affecting safe UAS operations are discussed.

Network Information
Related Authors (5)
Pengcheng Huang

23 papers, 635 citations

80% related
Mohit Jain

64 papers, 872 citations

74% related
Lothar Thiele

592 papers, 51.8K citations

73% related
Mitesh M. Khapra

133 papers, 2.6K citations

60% related
Shwetak N. Patel

265 papers, 11.9K citations

56% related
Performance
Metrics

Author's H-index: 20

No. of papers from the Author in previous years
YearPapers
202118
202012
20195
20189
20178
201613