Foundation Research Areas
Leading the way in AI research and innovation
Machine Learning & Deep Learning
Faculty in the department of AI are engaged in foundational research in machine learning (ML) and deep learning (DL), which are considered as the heart of AI and its success over the last decade. The branch of machine learning focuses on teaching computers to learn to perform tasks based on observed data, and deep learning forms a genre of methods in machine learning that has assumed a life of its own in recent years.
Faculty Involved
M. Vidyasagar
Sumohana S Channappayya
Vineeth N Balasubramanian
Srijith PK
J. Balasubramaniam
Subrahmanya Sastry Challa
Core Problems of Research
Deep learning architecture & training
The study, design and development of new deep learning architectures and efficient training methods have been crucial in driving the success of AI methods over the past few years. Faculty in the department have actively contributed to new model architectures, training methods and novel adaptations for various applications. Currently, the focus is on multimodal deep learning models integrating different modalities, like vision and language, to create more holistic solutions.
Generative AI
Generative AI has captured the fascination of the world with its ability to assimilate information, interact, assist, and create new data. The department’s research focuses on developing generative models through methods like generative adversarial networks, variational autoencoders, normalizing flows, and more recently, diffusion models. Our past and ongoing efforts include analyzing generative models, addressing their limitations, applying them to vision and language, and developing them in settings where data or label information is limited.
Optimization for Machine Learning
Optimization is the engine of modern machine learning and deep learning methods. Although an interdisciplinary field, optimization for machine learning methods, in particular, has attracted significant interest across the worldwide AI community, primarily because of its unique need to achieve generalization, i.e., provide effective solutions on data that has never been observed before. Faculty in the AI department work on methods for both convex and non-convex optimization. Going beyond, the efforts of the faculty have also included the study of the effectiveness of non-convex optimization methods for training deep learning models, both theoretically and empirically.
Bayesian learning
A fundamental approach to machine learning that is widely considered to be ideal, yet difficult to implement in its truest sense, is Bayesian learning. Founded on strong principles of Bayesian theory, this family of methods can accommodate the integration of prior knowledge as well as provide estimates of uncertainty for model predictions that can be very useful in risk-sensitive applications. AI department faculty have strong interest and expertise in Bayesian deep learning as well as Bayesian non-parametric models, where past and ongoing efforts include the development of efficient inference techniques for Bayesian deep learning models, Gaussian processes and Bayesian optimization. Our efforts have also included the application of Bayesian learning to problem domains in autonomous navigation, social media analysis and astrophysics.
Reinforcement learning and control theory
The need to perform sequential decision-making in many applications including gaming, navigation and robotics motivates the paradigm of reinforcement learning methods that learn policies towards an eventual goal, rather than an immediate prediction. The AI department includes faculty with strong theoretical interest in control theory and its direct relevance to reinforcement learning methods, as well as its applications to robot navigation and manipulation.
Learning & adaptation in evolving & data-scarce environments
While supervised learning, the harbinger of machine learning settings, has evolved and matured over the last decade, learning in evolving non-stationary environments where the data distribution may not necessarily be identically independently distributed (i.i.d) or may be scarce has grown to be an important need for AI research. This includes a wide variety of settings including continual learning, domain adaptation, few/zero-shot learning and domain generalization. Faculty in the department actively work on developing deep learning models for these settings, and study both theoretical and applied aspects of newer methods in this direction.
Algorithms for massive datasets
In recent years, more data has been produced than ever before in the history of mankind. On one hand, such volumes of data provide a goldmine for knowledge extraction, commercial value and policy making. On the other, building reliable models on such large volumes is cumbersome and, at times, intractable, despite access to high-end computational facilities. Faculty in the department work on addressing these challenges through scalable algorithmic solutions that are both efficient and effective using data sketches (summaries), statistical variance reduction techniques, as well as building noise/outlier robust algorithms to get more accurate predictions.
Neural differential equations
An emerging area of deep learning methods has been the intersection of neural network models with ordinary and partial differential equations. This field is in an early stage with significant promise, and our faculty focus on developing various differential equation based neural networks for machine learning problems. Our research herein has focused on designing continuous-depth deep learning models inspired from differential equations. Treating the computation of intermediate feature representation in deep learning models as a solution to differential equations has tremendous advantages in terms of model selection and reduced number of parameters.
Casual inference and learning
Global leaders in AI have vocally vouched the need for the design and development of Type 2 AI systems, that can not only predict but can also reason. A critical aspect of such reasoning-based AI systems is the ability to separate causal relationships in data from spurious correlations. Understanding such causal relationships in data has a wide range of applications in risk-sensitive and safety-critical applications ranging from healthcare to economics. Faculty in the AI department have focused on integrating principles of causality into the training of deep learning models, as well as providing causal perspectives in explaining the decisions of such models.
Computer Vision
Computer vision is a field of artificial intelligence (AI) that enables computers and systems to derive meaningful information from visual inputs such as digital images/videos and to suggest recommendations based on that information. Computer vision applications range in various areas, such as surveillance, healthcare, autonomous navigation, manufacturing, agriculture, defense, etc.
Faculty Involved
Konda Reddy Mopuri
Sumohana S Channappayya
Vineeth N Balasubramanian
Srijith PK
C Krishna Mohan
Areas of Application
Agriculture: Plant phenotyping using computer vision
Healthcare: automated analysis of choroidal images
Drone-based vision: Detection of objects from drone (low-resolution) imagery
Autonomous navigation: adding levels of autonomy to driving vehicles in developing countries, focusing on India
Human behavior understanding: Detection of emotions, human poses, gestures, human mobility aids etc.
Image and video quality assessment: aims to quantify and measure the perceptual quality of natural images and videos
Core Problems of Research
Adversarial robustness: developing vision models that perform consistently and reliably in the face of inputs that are deliberately crafted to mislead
Knowledge/Model Extraction attacks: retrieving useful and sensitive information from the deployed models and defending against such attacks
Learning from limited or weak supervision: making the learning less reliant on labeled datasets
Explainability: making the models more transparent and providing human-friendly explanations for their predictions
Learning from long-tailed (imbalanced) datasets: handling the imbalances in training data
Continual learning: focuses on the ability of a model to learn from and adapt to new information over time without forgetting what it has previously learned
Causal learning: understanding and modeling the cause-and-effect relationships between variables in a system as opposed to merely learning the correlations between them
Bayesian learning and uncertainty modelling: Developing algorithms capable of modelling uncertainty and detecting out-of-distribution data. This is important for high-risk vision applications such as autonomous driving and healthcare as they have to deal with out of distribution data.
Domain generalization: Developing algorithms that can generalize to unseen domains and deal with distribution shifts.
Robotics and Industry 4.0
From its inception, the primary goal of artificial intelligence (AI) was to forge entities capable of both cognitive reasoning and physical action reminiscent of human abilities. As the demands of the industrial revolution took hold, however, this unified goal diverged into two distinct paths. In one direction, the field of robotics evolved, prioritizing the development of reprogrammable machines adept at performing repetitive, precision tasks to cater to manufacturing needs. Simultaneously, the pursuit of human-like thinking continued within AI, focusing on cognitive modeling, decision-making processes, and intelligent problem-solving. Now, in the present era, these once-divergent paths have converged, giving rise to advanced AI systems that seamlessly integrate human-like cognition with autonomous physical capabilities. At the same time, application of AI techniques to specific robotic tasks which need not necessarily require human-like cognition is also expanding.
Faculty Involved
R. Prashanth Kumar
S. Suryakumar
Core Problems of Research
Reinforcement learning
Reinforcement learning (RL) in robotics enables robots to learn tasks through trial and error, guided by rewards or penalties. By iteratively optimizing their actions, robots can develop complex behaviors and decision-making processes. RL’s integration with robotics represents a leap towards adaptive, autonomous systems in dynamic environments. We develop neural controllers trained by reinforcement learning for legged robots such as quadruped robots, biped/humanoid robots, and multirotor UAVs. Specific focus is on systems which are underactuated and hard to control using conventional control systems.
Industry 4.0
With the advent of ‘smart’ and ‘intelligent’ machines, we are looking at the onset of the fourth industrial revolution, also referred to as Industry 4.0. The next-generation factories will depend on various technology verticals to make this happen. The research at IIT-H revolves around the enabling technologies for this, including additive manufacturing/3D printing and digital twins.
Natural Language Processing
Language is a ubiquitous mode of communication. Understanding such communications and acting on them to create an impact thus requires understanding the natural language content, and further processing based on this understanding. The NLP research at the Department of AI, IIT Hyderabad focuses on multiple problems from a foundational as well as functional perspective.
Faculty Involved
Manish Singh
Maunendra Desarkar
Sri Rama Murty Kodukula
Srijith PK
Core Problems of Research
Multilinguality
Multilingual Natural Language Processing (NLP) involves the development of NLP systems capable of handling multiple languages. Among the notable challenges in this domain is addressing low-resource languages (LRLs), characterized by limited available data or resources. We work towards the advancement of modeling frameworks that facilitate cross-lingual/multilingual transfer across various NLP tasks, specifically focusing on LRLs. Our primary objective is to create efficient and effective solutions that can be practically applied in real-world multilingual scenarios. We have achieved notable progress in enabling zero-shot technologies for several LRLs by employing methodologies such as language structure analysis, cross-lingual transfer learning, meta-learning, and other innovative approaches.
Conversational AI
The domain of Conversational AI tries to enable computers or automated agents to understand, process, and respond to human language in a natural and personalized way. It encompasses voice assistants, chatbots, and other AI-powered systems that can carry out meaningful conversations with users. In this research space, we have undertaken problems like these, including building chat-bot systems for traffic control, creating conversational agents, and developing dialogue state tracking systems, among others.
Controllable Text Generation
With the growing capabilities of Natural Language Generation models, it has become imperative to ensure that the generated text is clean and free from unwanted attributes such as racial bias, discrimination, hate etc. We work towards developing methods that can generate texts in a controlled manner, to avoid certain types of attributes (say, non-toxicity) in the generated text.
Temporal textual data
We work in the intersection of vision, language and other modalities, developing algorithms to process multi-modal data and problems involving vision and language such as image captioning, visual question answering, etc.
Exploring product reviews
Due to the massive volume of products and sellers available in e-commerce sites, users take help of customer reviews to make an informed decision about selecting the products and the sellers. Although reviews contain very valuable information, almost all e-commerce sites provide very simple navigation mechanisms for exploring the reviews. To aid review exploration we have built two review exploration systems, one based on product aspects and another one based on opinionated tags. We have also worked on detecting stance in comparative reviews using unsupervised algorithms. We have also built a context-based review recommendation system, where the context is a combination of user preference and product aspects.
Continual & life-long learning
We develop algorithms that can learn continuously across multiple NLP tasks arriving in some sequence and transfer knowledge across them for improved learning and performance from limited data.
Multi-modal learning
We work in the intersection of vision, language and other modalities, developing algorithms to process multi-modal data and problems involving vision and language such as image captioning, visual question answering, etc.
AI and Ethics
AI algorithms are increasingly used to make decisions pertaining to societal and economic value. For instance, healthcare indicators data, bank credit scores, and criminal risk scores are increasingly used by algorithms to decide health premiums, loan decisions and bail decisions across the world. While the algorithmic decisions are more accurate, they are not inherently immune from ethical scrutiny. This vertical studies ways to formalize relevant parameters such as fairness, privacy and explainability and evaluate AI algorithms, models and AI-based decision frameworks. These studies include tradeoffs between efficiency and desired social parameters, interventions, suggestions and guidelines for practitioners to quantify and mitigate problems with AI-based decision-making frameworks.
Faculty Involved
Ganesh Ghalme
Core Problems of Research
Fairness in AI
The fairness requirements encapsulate how AI-based decisions treat individuals or groups of society. Some of the relevant research questions we ask are: Are the marginalized groups given sufficient representation? Are AI-based decisions discriminatory towards one group/community? Are similar individuals treated similarly? If there are biases, where do they emerge from and is there a way to intervene and mitigate bias and what is the cost of doing so in terms of loss in efficiency? Finally, we seek to propose socially aware AI algorithms.
Applications: Recommendation systems, crowdsourcing, banking, sequential resource allocation, sponsored search, internet content creation.
Explainability in AI
AI-powered decision frameworks often appear as black-box solutions. A complex machine learning or deep learning/neural network-based models power these decisions with the help of a large amount of data. These decisions, often, are not self-explanatory. For instance, an ML model is often unable to explain why a certain fraction of individuals were denied loan or in the education sector denied admission while others were accepted. Explainable AI strives to seek such explanations from ML models and provide recourse i.e. suggestions to applicants on what should they do to improve.
Applications: Healthcare, Banking and insurance, education
Strategic AI
AI algorithms, for all their efficiency and speed, are still vulnerable to strategic manipulation. As most AI-based algorithms work on score-based systems (such as likes/shares in online content, health parameters and CIBIL scores in healthcare and banking, QS scores in institute rankings), it is only natural that strategic players ‘game’ the system by performing well on these parameters without improving on the true underlying quality. There have been instances where AI algorithms have been manipulated by individuals to get a loan (by increasing credit scores, for instance, by buying multiple credit cards) or lower health premiums etc. How can we make our AI algorithms robust to manipulation?
Application: Banking, insurance, education, health care, and social networks
Security and privacy
The leakage of private information and its subsequent exploitation is a major concern when one uses human data in sectors such as healthcare and banking. While one needs data to be able to make more efficient models, one should also ensure the privacy and security of an individual’s confidentiality.
Applications: Banking, online platforms