Steps Towards Networked Autonomy: Spatial Perception, Certifiable and Adaptive Learning Models, and Time Sensitive Information Exchange

Talk by Dr. Rajat Talak, MIT
1st December 2023, 5:00pm to 6:00pm IST
[Online, Open to All]

Abstract:

Networked Autonomy aspires to build a team of robots that can seamlessly communicate, perceive, and interact with humans and the environment. The ultimate tool for humankind is such a networked autonomous system, which humans can direct, for example, to (i) organize inventory in a hospital and manage patients; (ii) help cook, clean floors, wash dishes and laundry at home; (iii) assemble prefabricated parts, ferry construction material, plaster, and paint walls, install wirings and lighting, at a construction site; (iv) carry out system checks and repairs, setup experiments, and perform routine cleaning operations, on the international space station; (v) and even enable space exploration and asteroid mining. This is not possible today! 

My work has focused on building the next generation of robot perception and communication algorithms that can enable networked autonomy. For perception, robots need to build an actionable scene representation in real-time that is scalable for long-horizon autonomy. With deep-learning models (and their failure-prone behavior) finding their place in the robot perception system, it becomes necessary to develop mechanisms that can certify the correct workings of a model, and adapt the model, using real and unannotated data, to again, make it work to its desired specifications. Finally, networked autonomous systems operating in a constantly changing environment need to exchange time-sensitive information. I will present hierarchical 3D scene graphs as an actionable representation that can be constructed in real-time, from an image stream, and show that it is scalable for memory and computing. I will discuss our ongoing effort to develop certifiable and adaptive mechanisms for deep-learning models, such as object pose estimation. I will show how we have managed to design a new set of algorithms that can ensure a fresh exchange of time-sensitive information.

 

Learning Rates and Loss functions: A Critical Tool for Optimization

Prof. Snehanshu Saha, BITS Goa
26th October 2023, 5:00pm to 6:00pm IST
[Venue: Online, Open to All]

Abstract:

Robustness to noise and performance are key factors to choosing loss functions in Deep Learning based Regression and Classification tasks. Rigorous mathematical investigation of learning rates used in back-propagation in shallow neural networks has become a necessity. This is because experimental evidence needs to be endorsed by a theoretical background. Such theory may be helpful in reducing the volume of experimental effort to accomplish desired results. We leveraged the functional property of  loss functions for these tasks. I will talk about a framework for an adaptive learning rate policy for several loss functions (including some lesser known ones) and evaluate its effectiveness for regression tasks. The framework is based on the theory of Lipschitz continuity, specifically utilizing the relationship between learning rate and Lipschitz constant of the loss function. Based on experimentation, we have found that the adaptive learning rate policy enables up to 5x-20x faster convergence compared to a constant learning rate policy. I will share these results as well.

 

Computational Thinking in Education: The Why, What and How

Mr. Vipul Shah, Tata Consultancy Services
10th October 2023, 5:30pm to 6:30pm IST
[Venue: LH1, IIT Goa campus]

Abstract:

Computing plays a bigger role in the lives of young people today than ever before. Yet, there is a widespread lack of knowledge of what constitutes the core of computing: is it just the use of word processors and spreadsheets or are there more fundamental principles that underlie this science? Digital Skills are becoming ubiquitous with Problem Solving, Critical Thinking and Analysis, Technology Design and Programming, Reasoning and ideation being identified as some of the top digital skills that are relevant for a person to be successful in the digital economy. It has been argued that not only is Computational Thinking a fundamental skill used by and useful for all but is a practice that is central to all sciences, not just computer science. So, what is computational thinking? What are different curricula and pedagogical approaches? In this talk, we will present the approach we have explored for grades 1-8 since 2016 and how we can extend the principles to higher education. I will also talk about the CT course offered by BSc programme in Programming and Data Science at IITM.
 
 

The impact of Generative AI in teaching, pedagogy and assessment. 

Prof Viraj Kumar, IISc Bangalore
6th October 2023,  5pm to 6pm IST
[Venue: LT3, Main building, IIT Goa campus]

Abstract:

This talk will address a pertinent topic that resonates with both educators and students. For educators, the discussion will revolve around adapting pedagogical approaches and assessment methods in light of the accessibility of powerful tools like Chat GPT and Copilot. We will explore the transformative potential of these tools in shaping teaching methods and evaluating student learning. Meanwhile, for students, the focus will be on harnessing the capabilities of Generative AI effectively and developing the critical ability to critique AI-generated code.
 
 

Problems from Data Analyses on the Ground

Prof. Rajesh Sundaresan, IISc Bangalore
29th September 2023, 2pm to 3:30pm IST
[Venue: In-Person, at PCCE Verna, Goa]

Abstract:

Over the past few years, our group has been working on a variety of operations research and data analysis problems related to water distribution, disaggregation of energy consumption, feedback intervention to shape energy consumption behaviour, IoT network design, computational epidemiology, etc. In this overview talk we will highlight some interesting problem abstractions and the role of computing in arriving at some solutions.
 
 

Learning to Simulate Millions of Agents

Ayush Chopra, MIT Media Lab
4th August 2023, 6pm to 7pm IST [Online]

Abstract:

Humanity is facing grand challenges at unprecedented rates, nearly everywhere, and at all levels. Many of these challenges: pandemics, financial market instability and disinformation in social media, are emergent phenomena that result from complex interactions between a large number of strategic agents. Agent-based modeling (ABM) helps simulate such complex systems by modeling the actions and interactions of individual agents contained within. Their utility depends upon the ability to recreate populations with great detail, efficiently calibrate to real-world data and analyze sensitivity of results. However, ABMs are conventionally slow to execute, difficult to scale to large populations and tough to calibrate. My research aims to alleviate these challenges by fundamentally rethinking ABMs in this era of AI. First, we introduce algorithms advancements that allow us to rapidly simulate and analyze interventions on million-scale real populations. Second, we present from case studies that evaluate prospective immunization policies and analyze impact of retrospective decisions of the COVID-19 pandemic. Finally, we introduce AgentTorch, a platform to simulate and analyze million-scale populations: inside the body, around us and beyond!
 
 

Deep Learning Challenges in Industrial Vision

Dr. Shitala Prasad, IIT Goa
17th February 2023, 4pm-5pm  [Online]

Abstract:

This talk will describe our work on 3D object recognition.  While 3D recognition using 2D multi-views is a well-studied problem, it is challenging if the objects are texture-less and are only differentiable by their shapes at certain viewpoints. Existing methods are based mostly on supervised learning, requiring a large number of labeled images per object. In our work we introduced a multi-loss view invariant stochastic prototype embedding to improve the recognition accuracy of novel objects at different viewpoints. 

 

Applications of Natural Language Processing

Dr. Aditya Joshi, Data scientist at SEEK, Melbourne
28th October 2022, 4pm-5pm  [Online]

Abstract:

Natural language processing (NLP) is a sub-branch of artificial intelligence that deals with computational approaches to process human (natural) language. The current interest in NLP rests on its ability to automate several language tasks that seem ‘intelligent’: from answering questions when one types a query in a search engine to helping a doctor search medical literature. This talk will explore sub-areas of NLP such as sentiment analysis, summarisation, question-answering, text classification and information extraction. I will then discuss how these sub-areas can be applied to biomedical, education-technology and work productivity domain.
 

 

Introduction to Research in Computing Systems

Dr. Sharad Sinha, IIT Goa
6th May 2022, 4pm-5pm  [Online]

Abstract:

This talk is focused on the following points:
1. What is a computing system? 
2. Why is it interesting to do research in computing systems? Examples of a few latest research results.
3. Which skills are needed to kick off your research in this area?
4. Publishing at top systems conferences and in top systems journals
5. Building your research team and career in computing systems research.
 

 

Proving Lower Bounds

Dr. Nutan Limaye, ITU Denmark
28th January 2022, 7:00pm-8:00pm [Online]

talk_poster_Nutan_Limaye

Abstract: If I asked you to represent the square-root of two as a fraction or if I asked you to name the largest prime number, then  possibly after a bit of thought, you will say “that is impossible!”. In mathematics, it is not enough to claim that something is impossible,  but you should also prove it. I will take a computational view towards proofs of impossibilities and introduce you to some concepts from a fascinating area of Theoretical Computer Science called Complexity Theory.

 

 

Introduction to Discrete-Event Simulation – A hands-on Tutorial

Dr. Neha Karanjkar, IIT Goa
26th November 2021, 4pm-5:30pm [Online]

tutorial_poster

Abstract: Simulation plays a critical role in the analysis, design and optimization of complex  systems in many areas. This tutorial presents an introduction to the simulation of Discrete-event systems (that is, systems in which the state is assumed to change at discrete time-instants only, as opposed to a system where the state evolves continuously over time). Computer networks, manufacturing systems, queueing systems, digital logic circuits, inventory systems are some areas where discrete-event simulation is widely used. This tutorial aims to provide an intuitive understanding of how discrete-event simulation works, using several examples. I will also provide a brief introduction to SimPy, a discrete-event simulation library in Python. The tutorial will assume basic familiarity with the Python language. Some sample programs and reference material will be shared with the attendees.

 

 

 

Current Research Directions in Fair Division

Talk by Dr. Neeldhara Misra, IIT Gandhinagar 
22nd October 2021, 4pm-5pm [Online]

Talk Abstract: Have you ever wondered about how to cut a cake in a way that makes everyone happy? It’s easy to do when the cake is nice and uniform, but when different people like different parts of the cake, dividing it  fairly is a trickier issue. The cake is a useful abstraction for various real-world problems, ranging from land to water to compute resources on a machine, and so on. I will introduce the problem of fair division with an  emphasis on the setting of indivisible items and talk about some open problems in the area. 

 

 

Machine Learning for Humans

Talk by Prof. Ashwin Srinivasan, BITS Goa
17th September 2021, 4pm-5pm [Online]

Talk Abstract: This talk will not be about AI, but about IA. That is, it will be about some of the machine learning  issues that are unavoidable in the design of Intelligent Assistants. I envisage such tools as working collaboratively with people in Sciences, Engineering, Business and the Arts. The specific role for machine learning I have in mind is not that of developing interfaces to data, but to assist actively in the creative aspects of human endeavor in these areas. I will focus on hypothesis formation in Science as an example. For machine learning techniques to assist meaningfully in this activity, I think at least two kinds of abilities are necessary: (1) ML tools should be able to include, as input, what is known already to the people they are assisting; and (2) ML tools should be able to communicate, as output, what they have discovered, in a form that is understandable to the people they are assisting. In this talk, I will show some ways in which modern neural network learning could be used with old-fashioned logic-based learning to build intelligent assistants that can be used for man-machine collaboration in Science.

 

Probability and Statistics: Past, Present, and Future of AI

Talk by Dr. Satyanath Bhat, IIT Goa
20th August 2021, 4pm-5pm [Online]

Abstract: AI and ML is in the limelight of popular media. There is an unsaid understanding that AI/ML is a recent  phenomenon. However, researchers of yesteryear have laid the foundation to the area decades back. Probability and Statistics, in particular, heavily drives the modern AI/ ML algorithms. In this talk, we will first provide a probabilistic model for popular ML areas. Thereafter we will walk through applications of probability and statistics that are in wide use today.

ACM-W Celebration (Sept 2016)

https://goa.acm.org/acmwcelebration

ACM India, ACM-W India, GOA University along with ORACLE Academy jointly organized ACM-W Celebrations and the First ACM India National-level Hackathon for Women in Computing on 23rd and 24th of September 2016.

Annual Event  (Feb 2015)

http://goa.acm.org/annualevent2015 

ACM India’s annual event for the year 2015 was held at Birla Institute of Technology Pilani, K K Birla Goa Campus, from 5–7 Feb 2015.