Day 1 Hall 3

In the book Predictive Machines, Prof. Ajay Agarwal, explains the economics of AI as centered around bringing down the cost of ‘predictions’. In healthcare, there are several critical patient-centric questions that can be best formulated as predictions – predict which patient is at higher risk for a certain disease, predict which patient would benefit from a novel treatment, predict which patient would be a non-responder, predict which patient will not be adherent to treatment. These patient prediction problems offer a rich testing ground to evaluate a range of AI paradigms – from foundational to emerging and the value they can provide to address these high impact patient-centric questions.

In our ZS AI lab, we focused on a particularly complex version of these patient prediction problems – identifying patients who are suffering from rare conditions (affecting <1 in 2000 people) but are currently either undiagnosed or misdiagnosed. During the talk, we will motivate why this is both a hard and an important problem to address. We will highlight a series of systematic experiments that our ZS AI lab scientists conducted to tackle the key challenges underpinning this problem including noisy labels, high data cardinality, and the temporal patient journey dynamics. We will share the meta-learnings from our research - in particular, our findings on the relevance of semi-supervised and generative models. Given the complexity of both the central prediction problem and the overall healthcare domain, we are continuing to enhance our approach along a few key dimensions – the core prediction algorithms, technologies to handle data complexities, and improving model trustworthiness to drive broader adoption with healthcare stakeholders. We will share more on these forward looking aspects. Our broader goal at ZS is to improve health outcomes for all. Realizing this goal requires the best minds to come together, share, discuss, and collaborate. During the session we would, therefore, love to hear your thoughts and experiences with analogous problems and in adjacent fields. Please join us at the session – a small step towards the broader goal of moving from the current state of ‘sick-care’ to truly healthcare.

Over the past 24 months, companies in the oil and gas supply chain have been struggling with supply chain uncertainties and logistics. Challenging external forces are not new to oil and gas marketing companies, but the pace that it happened in last couple of years have forced them to adapt with greater speed and agility to address the current landscape and prepare for a new market reality. This talk covers how oil companies can conduct accurate forecasting and find out ways to fulfill the same using linear programming.

Travel e-commerce data science space is complex. While there are many success stories, the offline metric improvement/lift and online business metrics may not align in many experiments. The offline model development and deployment strategy suffices in majority of cases. But one must monitor concept drifts, data drifts, remove training data sample selection biases, remove target leakages, minimize omitted variable biases and handle other aspects in applied data science world. Also, on many occasions there may not be sufficient data points for every decision, to learn with high degree of confidence which decision is better or worse to recommend. Considering these factors, we found reinforcement learning or contextual bandits useful. Combine these with online/offline batch learning, and deep learning representation models. At MakeMyTrip-GoIbibo group, we have built in-house platform to facilitate data scientists to quickly launch such experiments and move towards continual learning. This talk is about the problems, platform, use cases/examples we solved upon with this framework, glimpses into a few bandit model forms, off-policy evaluation metrics we use, and also how we combined these models with multi-objective decision methods such as pareto frontiers.

Artificial intelligence promises to change virtually every aspect of our jobs and personal lives. Already, its influence is undeniable in most aspects of our lives. In the excitement over AI (and the constant debate about its negatives), people often forget that a human element remains critical to technological success. Yes, machines will play a greater role in the economy of tomorrow, and yes, they may replace select jobs that humans currently perform—but people will still be critical to the success of AI

This talk will address the advancements in the field of AI in the next decade, its interaction with humans and how this delicate balance needs to be addressed.

AI and Cloud Computing have transformed our lives drastically. Every day, people use digital assistants like Siri, Google Home, and Amazon’s Alexa, which execute spoken commands such as playing music on a connected speaker, buying a product online, etc. However, AI algorithms have one major drawback- they rely on cloud servers to perform computations and can’t power mobile phones, computers and other devices whose applications use AI. Hence AI has limited relevance in areas of sparse connectivity. Another issue caused by AI applications hosted in the cloud is privacy. There are scenarios in which sending personal data to the cloud for processing by ML applications might not be a possibility. Hence, doing ML inference on edge then becomes important.

In this session Gunnvant will talk about the edge ML ecosystem that is in place today and is accessible to tinkerers and hobbyists. A discussion on how anyone can get started with writing AI applications on edge and embedded devices will be undertaken. The session will also focus on hardware options that are in place, the software ecosystem that exists, the capabilities and limitations of contemporary stack.

Climate change is a globally relevant, urgent, and multi-faceted issue heavily impacted by energy policy and infrastructure. Addressing climate change involves mitigation (i.e. mitigating greenhouse gas emissions) and adaptation (i.e. preparing for unavoidable consequences). Mitigation of GHG emissions requires changes to electricity systems, transportation, buildings, industry, and land use.

In this Tech Talk we will be estimating the building energy consumption using various libraries of Intel® oneAPI Toolkits(Intel® AI Analytics Toolkit ) like

Intel® Distribution for Python* including highly-optimized scikit-learn and XGBoost libraries
Intel® Optimization for TensorFlow*
Intel® Optimization of Modin* (available through Anaconda* only).
Demonstration of running the notebook using Intel DevCloud .

The demo will also showcase how Intel oneAPI is six to seven times faster than the other frameworks