Handing higher-dimensional data poses a familiar challenge for most machine learning applications. In this talk, we discuss traditional regression models that induce shrinkage, such as LASSO, and their Bayesian interpretation. We then look at spike and slab priors that have become popular for high-dimensional Bayesian modeling. Lastly, we see how to use spike and slab priors to model sparsity in deep neural networks, computation challenges, and possible solutions.
AI-based technologies are slowly becoming an indispensable part of our lives. However, our interactions with these technologies have been limited due to their limited understanding of human behavior and common sense knowledge about the real world. In this talk, I will discuss the research we have been pursuing to integrate machines into human society. The talk is divided into two parts. Emotions or Affect are an integral part of human behavior and guide the decision-making process. To develop intelligent systems that interact naturally with humans, it is essential that they understand human affect. In the first part, I will talk about our latest work on developing models for understanding human Affect, its causes, and how it guides human decision-making. Many of the existing AI-based systems need to gain more understanding of common-sense knowledge about everyday activities that humans perform seamlessly. It is not easy to impart this knowledge about mundane activities to machines. In the second part of the talk, I will discuss ScriptWorld, an environment we have created to enable the development of agents that learn common-day activities from scratch via interactions with the environment. I will outline some of the RL agents we developed to learn about common-day activities like making coffee, planting a tree, fixing a bike, etc.
The talk will discuss the exploding area of intersection between econometrics and machine learning. First, I will discuss the trajectories that these two disciplines of econometrics and Machine Learning (ML) have traditionally taken and discuss their goals, settings and approaches. Second, I will outline the area of intersection between the ML literature and, econometrics literature. Here, I will discuss several streams in the literature that have emerged within this intersection. These include using Big Data in economics and policy-making, predictive modelling for socio-economic outcomes, supervised and unsupervised learning methods used in economics. Third, I will discuss some newly developed methods in the intersection of econometrics and ML which typically improve upon the performances of methods in traditional econometrics and traditional Machine Learning.
We propose and study a new privacy definition, termed Probably Approximately Correct (PAC) Privacy. PAC Privacy characterizes the information-theoretic hardness to recover sensitive data given arbitrary information disclosure/leakage during/after any processing. Unlike the classic cryptographic definition and Differential Privacy (DP), which consider the adversarial (input-independent) worst case, PAC Privacy is a simulatable metric that quantifies the instance-based impossibility of inference. A fully automatic analysis and proof generation framework is proposed: security parameters can be produced with arbitrarily high confidence via Monte-Carlo simulation for any black-box data processing oracle. On the utility side, the magnitude of (necessary) perturbation required in PAC Privacy is not lower bounded by Θ(√d) for a d-dimensional release but could be O(1) for many practical data processing tasks, which is in contrast to the input-independent worst-case information-theoretic lower bound. We discuss applications of PAC Privacy to statistical data processing tasks.
Accept-reject based Markov chain Monte Carlo (MCMC) algorithms have traditionally utilised acceptance probabilities that can be explicitly written as a function of the ratio of the target density at the two contested points. This feature is rendered almost useless in Bayesian posteriors with unknown functional forms. We introduce a new family of MCMC acceptance probabilities that has the distinguishing feature of not being a function of the ratio of the target density at the two points. We present a stable Bernoulli factory that generates events within this class of acceptance probabilities. The efficiency of our methods rely on obtaining reasonable local upper or lower bounds on the target density and we present an application of MCMC on constrained spaces where this is reasonable.
Drawings are a powerful cognitive technology for creating external representations of thoughts. Can they be used as an alternative or complementary to natural language for cross-modal tasks? We have investigated this in some of our recent works. This talk will be around this theme. Specifically, I will talk about sketch+text for Image Retrieval, sketch-guided object localization, and sketch-guided image inpainting works. Toward the end, I shall conclude the talk with open topics in this area.
Dialogue systems or Conversational AI agents are becoming increasingly popular. These adoptions are mostly backed by promising potentials and commercial values of such systems. With the advancements in Large Language Models (LLMs) and related research, the quality of the generated responses in dialogue systems has been further enhanced. This is in turn increasing the potential of such systems. However, there are many scenarios where organizations are hesitant in adopting such dialogue systems. This is mainly due to the black-box nature of generating the responses. For any dialogue system to be effective, its responses should be explainable. Without explainability, the confidence in the generated responses is low. Poor quality responses with a lack of explanation can hit user satisfaction and can hurt business and/or relations. In this talk, we will focus on having explainability in dialogue systems - from the perspectives of both modelling and evaluation.
We propose a method of quantum machine learning called quantum mutual information neural estimation (QMINE) for estimating von Neumann entropy and quantum mutual information, which are fundamental properties in quantum information theory. The QMINE proposed here basically utilizes a technique of quantum neural networks (QNNs), to minimize a loss function that deter- mines the von Neumann entropy, and thus quantum mutual information, which is believed more powerful to process quantum datasets than conventional neural networks due to quantum superposi- tion and entanglement. To create a precise loss function, we propose a quantum Donsker-Varadhan representation (QDVR), which is a quantum analog of the classical Donsker-Varadhan represen- tation. By exploiting a parameter shift rule on parameterized quantum circuits, we can efficiently implement and optimize the QNN and estimate the quantum entropies using the QMINE technique. Furthermore, numerical observations support our predictions of QDVR and demonstrate the good performance of QMINE.
The need for explainability of Deep Neural Network (DNN) models and the development of AI systems that can fundamentally reason has exponentially increased in recent years, especially with the increasing use of AI/ML models in risk-sensitive and safety-critical applications. Causal reasoning helps identify input variables that cause a certain prediction, rather than merely be correlated, and thus provide useful explanations in practice. Similarly, focusing on causal input-output relationships can help a DNN model generalize to out-of-distribution samples better, where spurious correlations in training data may otherwise mislead a model. This talk will introduce the growing field of explainable AI, summarize existing efforts and focus on one important aspect of causality in DNN models -- the notion of causal attributions between input and output variables of the model. We will do this from two perspectives -- firstly, we will study how one can "deduce" what causal input-output attributions an already-trained DNN model has learned, and provide an efficient mechanism to compute such causal attributions (based on our work published at ICML 2019). Secondly, we will explore the complementary side of this problem on how one can "induce" known prior causal information into DNN models during the training process itself (based on our work published at ICML 2022) . Both of these efforts are derived by a first-principles approach to integrating causal principles into DNN models, and can have significant implications on practice in real-world applications.
Last three to four decades have seen tremendous advancements in the fields of science and technology. These advancements encompassed a variety of areas like communications, information technology, medicine, social media, environment, governance and many more. Concomitant with these advances, new challenges related to the ethical aspects have come up and people, in general, have found themselves grappling with ethical dilemmas, some of them totally unforeseen. This talk will try to touch upon the various ethical dilemmas that confront today’s researcher in the fields of science and technology. Recent endeavours which have been taken up by the University Grants Commission with specific reference to PhD students will be elaborated.
In today's fast-paced world where instant gratification is the norm, PhD candidates may wish to fast-track their PhD. Unfortunately, there exists no such lawful recipe that can make the path of PhD aspirants very smooth, without any obstacles. Pursuit of a PhD works on the principle of no pain no gain. It is important to recognize early that they are in for a long haul and prepare themselves to face the challenges both emotionally and financially. The talk will highlight some of the challenges faced by the PhD candidates and emphasize upon some of the life skills and management techniques that they can learn side by side to their usual technical and scientific expertise.
Despite their success, the performance of neural networks has been shown to be brittle to mismatch between train and test distributions. Previous works have hypothesized that this brittleness is caused because deep networks rely only on simple features of the input (such as background or texture of images) to make decisions, while completely ignoring complex features. Surprisingly, we find that the features learnt by network’s backbone are sufficient for out of distribution generalization, however, the final classifier layer trained using ERM does not use these features optimally for the same. We posit two reasons for this:
Over the last decade, there has been an explosive increase in AI/ML applications—all thanks to powerful computational resources like GPUs and TPUs. Recent trends suggest that computational requirements of AI, ML, and Graphics applications are briskly outpacing the supply. Innovations are necessary for the semiconductor industry to keep up the pace. All such applications are data intensive; hence, conventional technology of moving the data from the memory to the core limits performance. However, researchers have proven we can fill this gap by introducing near-memory and in-memory computation. This talk will cover the basics of such computing resources and opportunities to revive, research, and develop such systems.
Speech and language technologies have gained widespread acceptance worldwide in recent years. With new technology users rapidly growing in India, it is an important challenge to build inclusive technologies that cater to these new users. There are notable challenges towards achieving this goal largely owing to India's rich linguistic diversity. Multilinguality among Indian users contributes to large variations in speech accents that adversely affect speech recognition performance, even on resource-rich languages like English. User in multilingual societies like India also make frequent user of code-switching, i.e. mixing multiple languages while communicating, that further complicates user-machine interactions. In this talk, we will look at some of our recent work that aims to address these important challenges and outline a vision for widespread adoption of speech and language technologies in India.
Introduction to GPU ,Docker and Containers ,Data Preprocessing using RAPIDS ,Model Training and Optimizations ,Introduction to TensorRT ,Introduction to Triton Inference Server
In the context of global climate change that threatens human civilization in multiple ways, climate sciences have become a crucial topic of research. The parallel progress in the field of Machine Learning and the availability of huge volumes of climate-related data from multiple sources has sparked off interest in using data-driven approaches to answer crucial questions in climate science. Such questions involve short-term weather forecasts, long-term climate simulations and discovery of new laws and relations. Recent research has shown that it is possible to use models and algorithms related to Machine Learning and Deep Learning to answer more such questions successfully than earlier. In this talk, we will discuss several such applications of Machine Learning and Data Science in the domain of Climate Sciences.
Dr. Arjun Jain will give an in-depth overview of the inner workings of autonomous agents, including their history of development, key modules that enable autonomy, and technologies that empower them, focusing on perception, localization, prediction, planning and control. He will also discuss current limitations of autonomous systems and propose solutions to overcome them, and also showcase research and advancements in true autonomy for GPS-denied environments at UAVIO labs.
Design thinking is a problem-solving approach that is used to develop innovative and user-centered solutions. It is a process that helps individuals and teams to understand the needs and perspectives of users, and to develop solutions that meet those needs in a creative and effective way.
The prototype selection problem aims to learn a sparse distribution of a source set such that it best matches a different target set. The applications of this problem include target subset selection, data summarization, and clustering, to name a few. In this talk, we present efficient algorithms for the prototype selection problem. Using the optimal transport theory, the proposed optimization formulation is an instance of submodular maximization, and therefore, we propose a greedy algorithm with simple updates. We further make use of the bandit setup to reduce the computations. The presentation is based on the papers [1, 2].