We present Samanantar, the largest publicly available parallel corpora collection for Indic languages. The collection contains a total of 46.9 million sentence pairs between English and 11 Indic languages (from two language families). In particular, we compile 12.4 million sentence pairs from existing, publicly-available parallel corpora, and we additionally mine 34.6 million sentence pairs from the web, resulting in a 2.8X increase in publicly available sentence pairs. We mine the parallel sentences from the web by combining many corpora, tools, and methods. In particular, we use (a) web-crawled monolingual corpora, (b) document OCR for extracting sentences from scanned documents (c) multilingual representation models for aligning sentences, and (d) approximate nearest neighbor search for searching in a large collection of sentences. Human evaluation of samples from the newly mined corpora validate the high quality of the parallel sentences across 11 language pairs. Further, we extracted 82.7 million sentence pairs between all 55 Indic language pairs from the English-centric parallel corpus using English as the pivot language. We trained multilingual NMT models spanning all these languages on Samanantar and compared with other baselines and previously reported results on publicly available benchmarks. Our models outperform existing models on these benchmarks, establishing the utility of Samanantar. Our data and models will be available publicly and we hope they will help advance research in Indic NMT and multilingual NLP for Indic languages.

Human Centred XAI focuses not only on conveying information regarding an AI's internal reasoning process, but also considering that this information needs to be consumed by people and therefore we must generate explanations built on cognitive theories which take into account human situation awareness models. I will give a few examples from my work whereby cognitive inspired explanations were used to mitigate automation bias and increase user trust and reliance, and talk about the importance of co-design and a human-centred approach.

Traditional biology generally looks at only a few aspects of an organism at a time and attempts to molecularly dissect diseases and study them part by part with the hope that the sum of knowledge of parts would help explain the operation of the whole. Rarely has this been a successful strategy to understand the causes and cures for complex diseases. The motivation for a systems based approach to disease understanding aims to understand how large numbers of interrelated health variables, gene expression profiling, its cellular architecture and microenvironment, as seen in its histological image features, its 3 dimensional tissue architecture and vascularization, as seen in dynamic contrast enhanced (DCE) MRI, and its metabolic features, as seen by Magnetic Resonance Spectroscopy (MRS) or Positron Emission Tomography (PET), result in emergence of definable phenotypes. At the Center for Computational Imaging and Personalized Diagnostics (CCIPD) at Case Western Reserve University, we have been developing computerized knowledge alignment, representation, and fusion tools for integrating and correlating heterogeneous biological data spanning different spatial and temporal scales, modalities, and functionalities. These tools include computerized feature analysis methods for extracting subvisual attributes for characterizing disease appearance and behavior on radiographic (radiomics) and digitized pathology images (pathomics). In this talk I will discuss the development work in CCIPD on new radiomic and pathomic approaches for capturing intra-tumoral heterogeneity and modeling tumor appearance. I will also focus my talk on how these radiomic and pathomic approaches can be applied to predicting disease outcome, recurrence, progression and response to therapy in the context of prostate, brain, rectal, oropharyngeal, and lung cancers. Additionally I will also discuss some recent work on looking at use of pathomics in the context of racial health disparity and creation of more precise and tailored prognostic and response prediction models

Explanations have been studied in the AI literature under multiple settings and guises, and have become particularly popular with the recent interest in explainable AI systems. However, like the proverbial blind men trying to describe their perspective of the elephant, the perspectives and prescriptions tend to be quite disparate and incomplete. In this talk, I will argue that explainable behavior and explanations are best understood in terms of the mental models--especially those the AI system constructs of the behavior and expectations of the human in the loop. The stark surface differences observed among the subcommunities of the Explainable AI universe can then be understood in terms of additional assumptions/restrictions placed by different sub-communities (often done without any explicit admission). I will then talk about a spectrum of challenges in generating explanations, especially in cases where the AI systems reason over representations that are inscrutable to humans. I will argue that AI systems will need to learn and use symbolic models that are comprehensible to humans to explain their decisions--even if they choose not to use such models for their internal reasoning.

Despite notable successes on several carefully controlled benchmarks, current machine learning (ML) systems are remarkably brittle, raising serious concerns about their deployment in safety-critical applications like self-driving cars and predictive healthcare. In this talk, I discuss fundamental obstacles to building robust ML systems and develop principled approaches that form the foundations of robust ML. In particular, I will focus on the role of data and demonstrate the need to question common assumptions when improving robustness to (i) adversarial examples and (ii) spurious correlations. On the one hand, I will describe how and why naively using more data can surprisingly hurt performance in these settings. On the other hand, I will show that unlabeled data, when harnessed in the right fashion, is extremely beneficial and enables state-of-the-art robustness. In closing, I will discuss how to build on the foundations of robust ML and achieve wide-ranging robustness in various domains including natural language processing and vision.