CS Katha Barta

People

2021

Upcoming Talks


Past Talks

Humans (e.g. crowd workers) have the remarkable ability to solve different tasks by simply reading textual instructions that define the tasks and looking at a few examples. NLP models built with the conventional machine learning paradigm, however, often struggle to generalize across tasks (e.g., a question-answering system cannot solve classification tasks) despite training with lots of examples. A long-standing challenge in AI is to build a model that learns a new task by understanding the human-readable instructions that define it. To study this, we build NATURAL INSTRUCTIONS, a large-scale dataset of diverse tasks, their human-authored instructions, and instances. We adopt generative pre-trained language models to encode task-specific instructions along with input and generate task output. Our results indicate that the instruction encoding helps models achieve cross-task generalization. Additionally, the instruction paradigm improves the accessibility of NLP modules to non-expert users by enabling fast model design through plain text instructions of a new NLP task. Backed by extensive empirical analysis on GPT3, we observe important features for successful instructional prompts and propose several reframing techniques for model designers to create such prompts. Our results show that reframing notably improves few-shot learning performance; this is particularly important on large language models, such as GPT3 where tuning models or prompts on large datasets is not feasible.

Currently, smell design is an enigma bordering alchemy. In spite of many attempts at understanding how smell is brought about, we still don't have an objective way of telling what a molecule will smell like, given only its chemical structure. The talk proposes to design algorithms capable of predicting the smell of a molecule with the help of data science and machine learning. It has its challenges right from the data collection to the understanding of smell itself. I have been working on these lines to design algorithms and instruments to predict the smell/taste of a molecule using its physico-chemical and structural properties since last 5 years . The most exciting part of being in this research is its direct connection with basic science as well as technology.

Flavor and fragrance is a multi-billion dollar industry and acts as a source of employment to many individuals. In this talk I will discuss the algorithms developed which helped in understanding and improving the technology of perfume making along with the science of perfumery and hence the economy. Besides, this research has helped in accelerating the works to understand the basic phenomenon of olfaction.

Abstract - Volume visualization is a technique in computer graphics where a 3D scalar field is visualised to highlight internal structures of a volume. We will discuss existing volume visualization algorithms and see how these can be applied to medical volumes like CT and MRI. Without effective visualization, understanding such medical data is difficult. To this end we will discuss newer methods such as a graph based volume visualization algorithm developed at our research group. We will also look at applying AI based intelligence to transform medical imaging data to more a meaningful representation.
Abstract - Natural language processing is the science of teaching computers to interpret and process human language. There are still many challenging problems to solve in natural language. Recently, NLP technology has leapfrogged to exciting new levels with the application of deep learning, a form of neural network-based machine learning. Nevertheless, deep learning methods are achieving state-of-the-art results on some specific language problems.
The presentation covers the motivation of using deep learning approaches for natural language processing supplement with the below case studies.
  • Language Corpus Development
  • Machine Translation
  • Text Summarization
  • Fake News Detection
  • Language Detection
  • Operant Motive Classification
Abstract - Increasing number of patents and papers attest to the fact that smart materials-based smart structures/systems have a high potential of being applied in our day-to-day applications. The current scenario shows that the physics, chemistry, and engineering disciplines are the ones who have any exposure to the field of smart materials. This exposure has motivated the researchers to study the material to develop innovative devices to assist and improve human life. The vast list of smart systems includes examples from tiny Nitinol-based self-expandable coronary stents to Piezo-based structural health/condition monitoring of huge composite wind turbine blades and many more. However, these deployments are mostly limited to small scales. This talk presents an introduction to smart materials and recent innovative research in the field of smart materials-based systems (such as structural health monitoring (SHM) of composites using piezoceramic transducers and others) to faculty and students of the computer science department at NISER. While the physical smart material’s components act as limbs for sensing and actuation of the smart structures, computer systems act as the brain for efficient operations of these smart structures. However limited understanding, knowledge, and expertise of smart materials among CS have throttled the applicability and development of smart structures on a large scale. A harmonious collaboration between computer science and engineering researchers can lead to the development of integrated Next-gen smart systems for the betterment of humanity all around the globe. This talk will also briefly present other successful applications of smart materials for SHM, actuation, and sensing in the field of civil, energy, and medical industry, which can be helpful for the development of future smart cities. All these research and development activities for the future could fall short if the next generation of engineers and researchers do not think out of the box with the knowledge they have.
Abstract - One of the greatest evolution in the last few decades has been the internet and its application. As of today, our day-to-day life involves some usage of the internet and networking. How exactly networking has been evolving and what we do today in this area. What are the new areas where research and work going on currently? Real-life use case and how networking solving the problem. And Few sample problem statement areas for optimization.
Abstract - We address the problem of localizing an (unauthorized) transmitter using a distributed set of sensors. Our focus is on developing techniques that perform the transmitter localization in an efficient manner, wherein the efficiency is defined in terms of the number of sensors used to localize. Localization of unauthorized transmitters is an important problem which arises in many important applications, e.g., in patrolling of shared spectrum systems for any unauthorized users. Localization of transmitters is generally done based on observations from a deployed set of sensors with limited resources, thus it is imperative to design techniques that minimize the sensors’ energy resources. In this paper, we design greedy approximation algorithms for the optimization problem of selecting a given number of sensors in order to maximize an appropriately defined objective function of localization accuracy. The obvious greedy algorithm delivers a constant-factor approximation only for the special case of two hypotheses (potential locations). For the general case of multiple hypotheses, we design a greedy algorithm based on an appropriate auxiliary objective function—and show that it delivers a provably approximate solution for the general case. We develop techniques to significantly reduce the time complexity of the designed algorithms by incorporating certain observations and reasonable assumptions. We evaluate our techniques over multiple simulation platforms, including an indoor as well as an outdoor testbed, and demonstrate the effectiveness of our designed techniques—our techniques easily outperform prior and other approaches by up to 50-60% in large-scale simulations and up to 16% in small-scale testbeds.