Department of Data Science

Indian Institute of Technology Palakkad&nbsp;(IIT Palakkad or IIT PKD) is a public autonomous&nbsp;engineering&nbsp;and&nbsp;research&nbsp;institute located in&nbsp;Palakkad,&nbsp;Kerala. It was&nbsp;one of the five new&nbsp;IITs&nbsp;proposed in the&nbsp;2014 Union budget of India. The campus was inaugurated on 3 August 2015. The institute has 94 faculty, 978 students, and 63 non-teaching staffs.

IT Palakkad currently offers four-year Bachelor of Technology (B.Tech) and two years MS programmes in 8 different engineering disciplines including&nbsp;Computer Science, Electrical, Mechanical, Civil among others.

&nbsp;

IIT Palakkad

In recent years, Coarse Grain Reconfigurable Architecture (CGRA) accelerators have been increasingly deployed in Internet-of-Things (IoT) end nodes. A modern CGRA has to support and efficiently accelerate both integer and floating-point (FP) operations. In this paper, we propose an ultra-low-power tunable-precision CGRA architectural template, called TRANSprecision floating-point Programmable archItectuRE (TRANSPIRE), and its associated compilation flow supporting both integer and FP operations. TRANSPIRE employs transprecision computing and multiple Single Instruction Multiple Data (SIMD) to accelerate FP operations while boosting energy efficiency as well. Experimental results show that TRANSPIRE achieves a maximum of 10.06× performance gain and consumes 12.91× less energy w.r.t. a RISC-V based CPU with an enhanced ISA supporting SIMD-style vectorization and FP data-types, while executing applications for near-sensor computing and embedded machine learning, with an area overhead of 1.25× only. © 2020 EDAA.

World University root node

Proceedings of the 2020 Design, Automation and Test in Europe Conference and Exhibition, DATE 2020

TRANSPIRE: An energy-efficient TRANSprecision floating-point Programmable archItectuRE

Deep Neural Networks (DNN) are very effective in high performance applications such as computer vision, natural language processing and speech recognition. However, these networks are vulnerable to adversarial attacks that infuses perturbations in the input data which are imperceptible to human eyes. In this paper, we propose a novel decision-based targeted adversarial attack algorithm which exposes the vulnerability of the underlying DNN when implemented on a resource constrained computing edge. Experimental results show that the proposed model performs 4 seconds(s) faster on an average, in a single perturbed image generation than the state of the art RED-attack, while consuming 15% less time for the entire dataset. © 2020 IEEE.

Department of Computer Science and Engineering

IEEE Workshop on Signal Processing Systems, SiPS: Design and Implementation

Fast and Efficient Decision-Based Attack for Deep Neural Network on Edge

Supervised learning algorithms require sufficient amount of labeled training data for learning robust prediction models. The field of Transfer Learning (TL) (also known as knowledge transfer) deals with utilizing knowledge from data-rich auxiliary domains to learn a reliable predictor for the domain of interest. This chapter presents a condensed review of the shallow TL literature (prior to the deep learning era). The chapter motivates the need for TL using an application. After an informal introduction to TL, a categorization of TL approaches based on the characteristics of the domains is presented. Next, the different transfer settings along with the challenges in each setting are described. The TL frameworks are delineated using a generic optimization problem. The chapter also discusses a few real-world applications used for benchmarking experiments for each transfer setting. Finally, the chapter concludes with some unexplored avenues in the TL research. © Springer Nature Switzerland AG 2020.

Domain Adaptation in Computer Vision with Deep Learning

Shallow Domain Adaptation

Streaming codes offer reliable recovery under decoding-delay constraint $\tau $ , of packets transmitted over a burst-and-random-erasure channel. Prior rate-optimal code constructions had field size quadratic in $\tau $ and employed diagonal embedding of a scalar block code of length $n$ within the packet stream. It is shown here that staggered diagonal embedding (SDE) under which the $n$ code symbols are dispersed across a span of $N \ge n$ successive packets leads to a simpler, low-complexity construction of rate-optimal streaming codes having linear field size. The limits of the SDE approach under the restriction $N \leq \tau +1$ are explored. Some binary streaming codes that are rate-optimal under this restriction are identified. © 1997-2012 IEEE.

IEEE Communications Letters

Simple Streaming Codes for Reliable, Low-Latency Communication

Quantities appear in search queries in numerous forms: companies with annual revenue of at least 50 Mio USD, athletes who ran 200 m faster than 19.5 s, electric cars with range above 400 miles and so on. Processing such queries requires the understanding of numbers present in the query to capture the contextual information about the queried entities. Modern search engines and QA systems can handle queries that involve entities and types, but they often fail on properly interpreting quantities in queries and candidate answers when the specifics of the search condition (less than, above, etc.), the units of interest (seconds, miles, meters, etc.) and the context of the quantity matter (annual or quarterly revenue, etc.). In this paper, we present a search and QA system, called Qsearch, that can effectively answer advanced queries with quantity conditions. Our solution is based on a deep neural network for extracting quantity-centric tuples from text sources and a novel matching model to retrieve and rank answers from news articles and other web pages. Experiments demonstrate the effectiveness of Qsearch on benchmark queries collected by crowdsourcing.

Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

Qsearch: Answering Quantity Queries from Text

Linking and Mining Heterogeneous and Multi-view Data

Multi-View Data Completion

The uptake of solar power generation is on the rise. This necessitates more research into developing data-driven intelligent methods that can perform effective analytics over power generation data to inform strategies to improve solar power generation systems. In this paper, we consider the utility of time series representation learning for analytics over power generation data. WaRTEm, a representation learning method that focuses on learning time series representations that are invariant to local phase shifts, is the focus of our investigations in this paper. We identify two metadata attributes for power generation sequences, month and CellID, as attributes that embed useful notions of semantic similarity between time series sequences. We evaluate the effectiveness of WaRTEm representations, as against using the raw time series sequences, in alignment to the month and CellID labellings, using accuracy over 1NN retrieval as an evaluation framework. Through empirical evaluations, we identify that WaRTEm embeddings are consistently able to achieve better representations when evaluated on 1NN accuracy. We also identify some features of WaRTEm that are more suited for time series representation learning, which provides promising directions for future work. © 2019 IEEE.

2019 20th International Conference on Intelligent System Application to Power Systems, ISAP 2019

Time Series Representation Learning Applications for Power Analytics

We study the problem of designing error-correcting codes over channels with burst and random erasures, when a strict decoding delay constraint τ is in place. Badr et al. introduced a channel model wherein for any sliding-window of width w, at most one of the following erasures patterns are permissible; (i) a burst erasure of length ≤ b or (ii) a total of ≤ a random erasures. We present a rate-optimal code construction under this model, which covers all feasible channel and delay parameters. In contrast to existing rate-optimal code families which require a field-size at least as large as O\left( {\binom {\tau} {a} } \right), our construction needs a field-size quadratic in the decoding delay constraint. For some parameters, the construction can be over linear field-size. © 2019 IEEE.

IEEE International Symposium on Information Theory - Proceedings

A Quadratic Field-Size Rate-Optimal Streaming Code for Channels with Burst and Random Erasures

Transfer learning across heterogeneous feature spaces can, in general, be a very difficult problem in practice due to the heterogeneity of features and lack of correspondence between data points of different domains. In this paper, we present a novel supervised domain adaptation algorithm (SHDA-RF) that transfers knowledge from a data-rich source domain to a target domain with only few training instances. The proposed method makes use of random forests to identify pivot features that bridge the two domains. The key idea of the proposed feature transfer approach is that every path in a decision tree leading to a partition of the data is associated with a certain label distribution and the label distributions that appear both in the source and target random forest models can be used as pivots for bridging the two domains. This information is used to generate a sparse feature transformation matrix, which maps patterns from the source feature space to the target feature space. The target model is then retrained along with the projected source data. We conduct extensive experiments on diverse datasets of varying dimensions and sparsity to verify the superiority of the proposed approach over other baseline and state of the art transfer approaches.

Artificial Intelligence

Supervised heterogeneous feature transfer via random forests

Time series are ubiquitous in real world problems and computing distance between two time series is often required in several learning tasks. Computing similarity between time series by ignoring variations in speed or warping is often encountered and dynamic time warping (DTW) is the state of the art. However DTW is not applicable in algorithms which require kernel or vectors. In this paper, we propose a mechanism named WaRTEm to generate vector embeddings of time series such that distance measures in the embedding space exhibit resilience to warping. Therefore, WaRTEm is more widely applicable than DTW. WaRTEm is based on a twin auto-encoder architecture and a training strategy involving warping operators for generating warping resilient embeddings for time series datasets. We evaluate the performance of WaRTEm and observed more than 20% improvement over DTW in multiple real-world datasets.