Schedule

Introduction to the event, platform and organizers.

TBA

TBA

Artificial Intelligence (AI), especially the field of machine learning (ML), is transforming virtually all aspects of our lives. This talk describes our preliminary efforts at Google Research India to tackle problems arising in the Indian context and beyond. We start with examples of opportunities to apply ML to accelerate science and its applications, and some early results. We then describe a critical need and an opportunity to improve health outcomes globally at lower costs, and specific challenges in countries like India of an acute shortage of doctors. As examples, we describe a convolutional neural network based solution developed by our researchers for more effective screening of diabetic retinopathy, as well as ongoing efforts towards prevention of cardiovascular disease. We describe our AI for Social Good program through which we are addressing issues like public health, education and wildlife conservation, in partnership with NGOs and academic researchers. We describe challenges arising due to factors like diversity of languages and code mixing for Google products like Search and Assistant, which are being used daily by tens of millions of users in India, and our natural language processing research. Finally, we describe outstanding challenges associated with the overall field of AI itself, like safety, data privacy, and fairness, and our guiding principles and directions to address them.

In this talk we propose the DeepXML framework for deep extreme multi-label learning where the objective is to develop deep architectures for annotating data points with the most relevant subset of labels from an extremely large label set. We demonstrate that DeepXML can: (a) be used to analyse seemingly disparate deep extreme classifiers; (b) can lead to improvements in leading algorithms such as XML-CNN & MACH when they are recast in the proposed framework; and (c) can lead to a novel algorithm called Astec which can be up to 12% more accurate and up to 40x faster to train than the state-of-the-art for short text document classification. Finally, we show that when flighted on Bing, Astec can be used for personalized extreme classification for billions of users by handling billions of events per day, processing more than a hundred thousand events per second and leading to a significant improvement in key metrics as compared to state-of-the-art methods in production.

A Robot Scientist is a physically implemented robotic system that applies techniques from artificial intelligence to execute cycles of automated scientific experimentation. A Robot Scientist can automatically execute cycles of hypothesis formation, selection of efficient experiments to discriminate between hypotheses, execution of experiments using laboratory automation equipment, and analysis of results. The motivation for developing Robot Scientists is to both to better understand the scientific method, and to make scientific research more efficient. The Robot Scientist ‘Adam’ was the first machine to autonomously discover scientific knowledge: it formed and experimentally confirmed novel hypotheses. Adam worked in the domain of yeast functional genomics. The Robot Scientist ‘Eve’ was originally developed to automate early-stage drug development, with specific application to neglected tropical disease such as malaria, African sleeping sickness, etc. More recently my colleagues and I have adapted Eve to work on yeast systems biology, and cancer. We argue that it is likely that advances in AI and lab automation will drive the development of ever-smarter Robot Scientists. The Nobel Turing Challenge aims to develop ‘AI Scientists’: AI systems capable of making Nobel- quality scientific discoveries highly autonomously at a level comparable, and possibly superior, to the best human scientists by 2050. If this comes to pass it will transform our understanding of science and the Universe.

TBA

Many biological applications require the segmentation of cell bodies, membranes and nuclei from microscopy images. Deep learning has enabled great progress on this problem, but current methods are specialized for images that have large training datasets. Here we introduce a generalist, deep learning-based segmentation algorithm called Cellpose, which can very precisely segment a wide range of image types out-of-the-box and does not require model retraining or parameter adjustments. We trained Cellpose on a new dataset of highly-varied images of cells, containing over 70,000 segmented objects. To support community contributions to the training data, we developed software for manual labelling and for curation of the automated results, with optional direct upload to our data repository. Periodically retraining the model on the community-contributed data will ensure that Cellpose improves constantly.

Generative Adversarial Networks have been a popular component in the field of artificial intelligence in the past decade. In this talk, we will begin with a brief tutorial on how GANs work, and the various considerations involved while designing GAN architectures. We will then proceed to some of the more popular GAN architectures and discuss them from various perspectives including interpretability and ethics. Finally we shall discuss more recent developments on the use of GANs, including tackling real world problems.

Artificial Intelligence is now becoming increasingly intertwined with our social realities. This opens up multiple pathways of approaching the relationship between these two spheres: using AI to study the social sciences, and using the social sciences and critical theory to study AI. In this talk, he will be giving examples from my research conducted at the Knowledge Lab as well as other labs across the world which are doing pioneering work in this field. He will touch topics ranging from Cognitive Science and Sociology to Knowledge production and the systemic discrimination which AI can contribute to if we are not critically thinking about the issues at hand.

TBA

The volume, veracity and diversity of large chunks of data make it difficult for comprehension and processing within the limited computational infrastructure and human cognition. Essentially, the advent of big data is pushing Artificial Intelligence to draw increasingly empirical inferences while learning from reasonably "small" data relies on sound statistical and mathematical foundations. The debate is thus whether the AI community moves away from the path of "proof" based techniques while dealing with big data efficiently. In this talk, he'll touch upon the progression of AI from classical learning paradigms to deep learning and the implementation/resource challenges it poses. He'll briefly present the argument in favor of parsimony computing by discussing the trade off between CPU and GPU resource utilization. He shall argue in favor of conceptual AI by stressing on cheap, fast and reliable training of Neural Networks, one of the driving tools in predictive AI.

A serious challenge when finding influential actors in real-world social networks, to enable efficient community-wide interventions, is the lack of knowledge about the structure of the underlying network. Current state-of-the-art methods rely on hand-crafted sampling algorithms; these methods sample nodes and their neighbours in a carefully constructed order and choose opinion leaders from this discovered network to maximize influence spread in the (unknown) complete network. In this work, we propose a reinforcement learning framework to discover effective network sampling heuristics by leveraging automatically learnt node and graph representations that encode important structural properties of the network. At training time, the method identifies portions of the network such that the nodes selected from this sampled subgraph can effectively influence nodes in the complete network. The output of this training is a transferable, adaptive policy that identifies an effective sequence of nodes to query on unseen graphs. The success of this policy is underpinned by a set of careful choices for embedding local and global information about the graph, and providing appropriate reward signals during training. We experiment with real-world social networks from four different domains and show that the policies learned by our RL agent provide a 7-23% improvement over the current state-of-the-art method.

Recent advances in machine learning, specifically deep learning, have great potential in healthcare. In medical imaging, deep learning based approaches have been used in diagnosis, prognosis, and response assessment for applications in radiology, pathology, opthalmology, oncology and others. However, there are a number of challenges in developing robust and reliable models. Deep learning typically requires access to large, diverse and well-annotated datasets. These models can be brittle and may not generalize well. Often they are “black boxes” lacking the transparency necessary to build trust in them. They can encode and propagate biases. Using examples from radiology and opthalmology, we will discuss these challenges and consider approaches for mitigation. Finally, using COVID-19 as a use case, we will discuss how these algorithms can be used at an individual level for diagnosis and risk prediction, at a hospital system level for planning and resource management, at a regional level to map outbreaks and identify racial disparities.

TBA

Domains such as law, healthcare, and policy often involve highly consequential decisions which are predominantly made by human decision-makers. The growing availability of data pertaining to such decisions offers an unprecedented opportunity to develop ML models which can aid human decision-makers in making better decisions. However, the applicability of ML to the aforementioned settings is limited by certain fundamental challenges:

1. The aforementioned settings call for the design of models that account for fairness and interpretability. However, most of the existing ML models are primarily optimized for predictive accuracy and are not inherently fair or interpretable.

2. These settings are prone to missing counterfactuals problem i.e., the data only captures the outcomes of the decisions made by human decision-makers and not the counterfactuals.

3. The data available in these settings is often prone to a variety of selection biases.

In this talk, I will probe the aforementioned challenges in detail and discuss their implications. I will also provide a brief overview of the solutions designed to address some of these challenges.