The mmWave frequency modulated continuous wave (FMCW) radars provide excellent target detection and localization performance in harsh environments as well as low lighting conditions. They offer a resolution of less than 5 cm, a range detection range of hundreds of meters, and a velocity of up to 300 km/h. This talk focuses on the applications of these highly versatile sensors in automotive and industrial applications.

In recent years, the Internet of Things (IoT) is getting high attention among the research community due to its various advantages in the context of social and commercial issues such as cognitive health monitoring, smart agriculture, intelligent transportation, and industry 4.0. Wireless Sensor Networks are working as the backbone for IoT, where, the data is transmitted at the base station either directly or in a multi-hop manner. Direct data transmission over long distances consumes a large amount of energy for the individual device. On the other hand, multi-hop data transfer results in increased data transmission delay, more data interference, and reduced data throughput. In order to address these issues, a novel small-world phenomenon can be introduced into the network, leading to an optimal number of hops required for the data transmission. In this talk, we are going to discuss small-world characteristics and their advantages in improving the energy efficiency and quality-of-service (QoS) of IoT networks. We will discuss various methods to introduce small-world characteristics first, subsequently, machine learning frameworks will be explored for the development of small-world wireless sensor networks for IoT applications.

Abstract :

-----------

Recent developments in sparse optimization theory aim at providing certain classes of linear systems with sparse descriptions, which have the potential of being used for data-driven learning. The presentation aims at exploring the basic theory and applications of the sparsity driven learning methods.

Abstract :

-----------

Researchers are accumulating increasing evidence that machine learning models based on correlation are brittle. These models do not distinguish between correlation and causation. Accordingly, they provide limited insight and cannot reason about the effects of actions for decision making. The traditional fields of logical engineering, causal inference, and deep learning have struggled with learning causal models at scale for several decades, each approach facing substantial obstacles. In this talk, I present the emerging paradigm of Tsetlin machines. Tsetlin machines partially unify the latter paradigms through a fundamental shift from arithmetic-based to logic-based machine learning. Like logical engineering, a Tsetlin machine produces propositional/relational Horn clauses (logical rules). However, the logical expressions are robustly learnt using finite state machines in the form of Tsetlin automata. Tsetlin machines further handle uncertainty by using multiple clauses to signify confidence. In this way, Tsetlin machines introduce the concept of logically interpretable learning, where both the learned model and the learning process are easy to follow and explain. The paradigm has enabled competitive accuracy, scalability, memory footprint, inference speed, and energy consumption across diverse tasks. Recent progress deals with classification, convolution, regression, image analysis, natural language processing (NLP), and speech understanding. At the end of the talk, I cover the Tsetlin machine research horizon, addressing the unresolved learning challenge in logical engineering, the scaling challenge of probabilistic causal models, and the correlation-reliance of deep learning. The proposed research agenda merges: (1) recursive logical Horne clauses for modeling all computable functions; (2) causal Tsetlin machine learning for distilling causal mechanisms from data into causal Horn clauses; (3) large-scale probabilistic causal reasoning over sparse truth tables in Horn clause form.

Abstract :

-----------

BITS Pilani recently signed an MoU with the Telangana State Horticulture University to develop and deploy technologies to minimize the post-harvest loss of horticulture produce. With increasing population, while there is an urgent need to increase the production of fruits and vegetables, post-harvest losses mean decrease in production in real terms. In India these losses may amount to 30-40% of total production resulting in financial losses to the tune of hundreds of crores of rupees. In this talk, we present a framework that seeks to deploy technologies such as Sensors, IoT, Cloud and Machine Learning to reduce the losses and improve the productivity thereby saving the country of hundreds of crores of rupees and farmers from the financial burden.

Abstract :

-----------

This talk focuses on two families of metrics over measures, which are popular in Machine Learning: namely, the Mercer kernel based Maximum Mean Discrepancy (MMD) metrics and the Kantorovich-Wasserstein metrics.

In the first part of the talk we revise the relevant definitions, summarize basic properties, and discuss ML applications where the metrics lead to state-of-art performance. We then compare and contrast the two families side-by-side and note their complementary properties.

In the second half, we present our recent attempt to bridge the gap between the two by proposing a new family of metrics that seems to achieve the best of the both. We provide some theoretical insights and present empirical evidence to showcase the proposed metric's potential in ML applications.

Abstract :

-----------

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.

Abstract :

-----------

Humans learn continually throughout their lifespan by accumulating diverse knowledge and using the accumulated knowledge to efficiently learn new tasks without forgetting old tasks. Though deep learning models manage to achieve state-of-the-art performance in several Artificial intelligence tasks, in such continual learning situations and problem setups, they suffer from catastrophic forgetting where it forgets the knowledge gathered from previous tasks after learning a new task. A. deep learning model with a continual learning capability can be beneficial to several applications such as Autonomous driving, healthcare, chatbots etc. In this talk, we motivate and discuss continual learning in deep learning models and briefly mention the techniques to achieve it. We discuss the techniques that we proposed for continual learning (CL) in neural networks using hypernetworks and continual image generation techniques. We demonstrate the effectiveness of these techniques on several real world problems and setups.

Abstract :

-----------

In the recent past, deep learning architectures have been highly successful in signal processing applications, paving paths for end-to-end systems. It would be interesting to see how these deep architectures are related to the traditional signal processing (filter design) approaches. This talk offers a signal processing view of deep learning models. It is mainly aimed at beginners in understanding the parallels between both approaches. Finally, the application of deep architectures in speech enhancement will be discussed.

Abstract :

-----------

Combinatorial optimization is a problem that is widely studied in the domain of classical computation. Now, it is also receiving a lot of interest from a quantum standpoint. Many quantum algorithmic techniques are being studied, as well as applications. We will review these techniques and applications, including, if time permits, some work being done at IITH.

Abstract :

-----------

In this talk, I will focus on Anticipating Human Behavior.

First, I will talk about action anticipation. The ability to anticipate future actions of humans is useful in application areas such as automated driving, robot-assisted manufacturing, and smart homes. The problem of anticipating human actions is an inherently uncertain one. However, we can reduce this uncertainty if we have a sense of the goal that the actor is trying to achieve. present an action anticipation model that leverages goal information for the purpose of reducing uncertainty in future predictions. So, predicting the next action in the sequence becomes easier once we have an idea about the goal that guides the entire activity. We present an action anticipation model that uses goal information in an effective manner.

Second, I will talk about predicting human driving behavior in traffic especially at intersections as a large proportion of road accidents occur at intersections. Especially, in India where vehicles often ply very close to each other, it is essential to determine collision-prone vehicle behavior. Existing approaches only analyze driving behavior in lane-based driving. We learn the collision propensity of a vehicle from its interaction trajectory. Interaction trajectories encapsulate hundreds of interactions for every vehicle at an intersection. The interaction trajectories that match accident characteristics are labeled as unsafe while the rest are considered safe. We also introduce the first laneless traffic aerial surveillance dataset called SkyEye to demonstrate our results.

Abstract :

-----------

5G standards have stringent requirements with respect to handling very high data rates, low latency and supporting diverse services. One of the use cases of 5G interest is enhanced Mobile Broad Band (eMBB). Channel coding is employed in a communication system in order to mitigate the effects of channel noise and decode the transmitted messages even in the presence of errors. In this talk, we will briefly review two classes of codes used in 5G new radio standard, which are LDPC codes and Polar codes. LDPC codes are a class of codes which can achieve rates quite close to the channel capacity. These codes are characterized by parity check matrices which are sparse in nature. LDPC codes can be decoded using belief propagation algorithms on Tanner graphs. The belief propagation algorithm itself is highly efficient and scales very well with the block length of the code. LDPC codes are used for transmissions on the data channel, where the block lengths are of the order of multiple thousands. Polar codes are the first class of codes which are provably capacity achieving, for a range of channels. Polar codes are designed based a certain polarizing transform. These codes can be encoded using a recursive algorithm since the polarizing transform is a kronecker product of a certain 2 X 2 matrix. Also, decoding of polar codes is performed using successive cancellation technique. Polar codes are used for transmissions on the control channel, where the block length roughly ranges from 256 to 1024. Recently, there have been advances in designing codes and decoders based on deep learning and we will present important results in this area. Also, there have been developments with respect to Reed-Muller codes achieving the capacity of binary symmetric channel. We will also briefly discuss these results.

Abstract 1 :

-----------

Light-weight Convolutional Neural Networks (CNNs) are mobile friendly models that can provide inference without the need for any specialised hardware. These models can be very effective in point-of-care settings, where the detection of disease has to be performed in real-time. The talk will highlight few developments of Medical Imaging Group (MIG) at Indian Institute of Science, especially towards COVID19 management and diagnosis. Deployment of these lightweight networks on embedded platforms to show highly versatility as well as prove optimal performance in terms of being accurate will also be highlighted. The developed models having latency in the same order as other lightweight networks without compromising the accuracy will also be shown.

Abstract 2 :

-----------

There is a lot of similarity between self-driving cars and current AI adoption in medicine, especially in Radiology. Today, AI solutions for radiology are focused towards image enhancement. However, with the advent of innovation and the positive potential impact of AI augmentation, radiologists should expect to see more of automated diagnosis in the near future. The long-term vision for AI in medical imaging should stop at autonomous vehicle level 3 automation—the radiologist will still be in charge of the majority of cases, and AI can help them drive along the "highway" of simple use cases. The talk will focus on current challenges and list opportunities for the researchers.

Abstract :

-----------

Second-order optimization methods are among the most widely used optimization approaches for convex optimization problems, and have recently been used to optimize non-convex optimization problems such as deep learning models. The widely used second-order optimization methods such as quasi-Newton methods generally provide curvature information by approximating the Hessian using the secant equation. However, the secant equation becomes insipid in approximating the Newton step owing to its use of the first-order derivatives. In this study, we propose an approximate Newton sketch-based stochastic optimization algorithm for large-scale empirical risk minimization. Specifically, we compute a partial column Hessian of size (d × m) with m ≪ d randomly selected variables, then use the Nyström method to better approximate the full Hessian matrix. To further reduce the computational complexity per iteration, we directly compute the update step (∆w) without computing and storing the full Hessian or its inverse. We then integrate our approximated Hessian with stochastic gradient descent and stochastic variance-reduced gradient methods. The results of numerical experiments on both convex and non-convex functions show that the proposed approach was able to obtain a better approximation of Newton's method, exhibiting performance competitive with that of state-of-the-art first-order and stochastic quasi-Newton methods. Furthermore, we provide a theoretical convergence analysis for convex functions.

Abstract :

-----------

We will discuss various applications of machine learning in wireless networks for classification and localization. This includes using existing wireless networks like Wi-Fi for 3-dimensional (3D) indoor localization. State-of-the-art multi-building 3D indoor Wi-Fi localization schemes. A proposed CNN based scheme that performs significantly better than the state-of-the-art schemes based on openly available Wi-Fi datasets. Other works like machine learning for target classification by mmWave RADARs, vehicle detection using LTE signals, etc.