Day 2 - 1st OctoberThursday, 1 October 2020

A team of engineers worked tirelessly to build chatbots that could talk to us like a good friend. But this has led to an empire of fat chatbots that couldn’t deliver on their promises.

Brands and customers alike were left disappointed and frustrated. But now, a band of rebels is applying AI pragmatically to deliver a new hope for the future of customer service, where the empire of fat chatbots is replaced by lean and agile chatbots that truly bring customer support into the digital age.

Join this session with Helpshift founder and CEO, Abinash Tripathy and learn a pragmatic approach to AI and chatbots that delivers on its promises

The very fast and drastic changes in consumer behaviour, supply reliability and legal restrictions due to the Corona virusare a huge challenge for supply chains and retailers. The first reaction of most companies was to switch off automatic systems and give control to humans.

However, we show some examples that advanced automation and clever adaptable AI ordering and pricing algorithms led to better than human decisions even in these uncertain times and gave a competitive advantage in overcoming the crisis.

As ML model performance improves and transformational tools like Auto ML lower the barriers to ML adoption, high quality ML outputs will become ever more common. Yet without high quality product implementations, these modeling improvements will fall short of their potential impact or, worse, drive unintended outcomes. In this session, Will Pearce, Product Manager for Google’s internal Applied Machine Learning team, will explore the actions product leaders should take at every stage of the ML project lifecycle to build successful ML-powered products.

• In a data-led era those who can use machines algorithms to identify trends and insights from vast data gain the competitive advantage. How have the leaders done it, and what can we learn?
• What kinds of data are required to build a holistic picture of your market opportunities? What kinds of models are possible?
• How to go beyond individual AI applications to deliver holistic AI platforms that reach across every aspect of the business
• Real-world ways to leverage software and data for real-time identification of market opportunities & threats

Artificial Intelligence (AI) is increasingly being used in mission critical applications. AI has not only been successfully applied in development and delivery of customer facing products and services, but also to improve internal processes such as maintenance and skills management. However, there also have been spectacular failures with dramatic consequences for users’ health and accidents leading to loss of assets. Unfortunately, currently available methods of testing and quality assurance have thus far failed to appropriately address the black box nature of AI. Current methods are not even able to accurately quantify the risks associated with AI components.

The AI Quality Management System (AI-QMS) offers a transparent and objective framework to identify, manage, and mitigate risks associated with AI technologies. It defines an AI specific life cycle model for development and operation of AI based products and services, which covers both, AI technology and AI data. The AI-QMS guides through the selection of quality goals and assists monitoring by defining conditions of acceptability in response to common pitfalls at each life cycle phase. As a result, a quality mark can be achieved which elevates trustworthiness and minimizes uncertainty.

In this talk, Julian Sanchez will explore where human factors psychology intersects with Artificial Intelligence…on the farm. While humans are the ultimate sensors of the world we inhabit, AI systems can significantly augment our productivity by providing highly reliable decision support systems and semi-autonomous features – Julian will discuss how this is already happening on farms today.

AI’s dependence on open-source software technologies is heavy and widely known. Efforts to develop open-source AI tools and libraries have been widespread and have already made major impacts across all stages of the development lifecycle, including annotation tools, data management tools, core software libraries, experimentation and even deployment.

Interestingly these efforts have been contributed by stakeholders across the gamut, including academics, startups and established industry. This talk will discuss these tools, making an effort to understand certain key questions at the intersection of AI and open-source software: Is AI unique in its reliance on open-source? Are the open-source tools enabling AI or limiting it? Will the dependence on open-source tools create a fragility in maintaining them in the future? I will discuss these questions in the context of specific toolsets in the open-source AI community.

The future of work remains unpredictable and uncertain. More than ever before, business leaders need to remain confident that their operations can continue securely in the face of global or even regional crises, and while sections of the economy are slowly re-opening, cyber-attackers are ramping up their campaigns.    

As businesses look set to rely on cloud and SaaS tools for the long term, our digital environments are going to be more dynamic than ever. Yet organizations are finding themselves undergoing a delicate balancing act—each new work practice and technology that is introduced also brings unforeseen risk. Static, legacy approaches have become redundant, both unintelligent and ill-equipped to adapt.  

Organizations must rethink their approach to security ,and rely on new technologies like AI to achieve much-needed adaptability and resilience. Darktrace is the world leader in Cyber AI technology, and leverages unsupervised machine learning to seamlessly adapt and integrate into changing environments, and to detect and respond to attacks in the earliest moments.    

In the face of an uncertain present and future, Cyber AI enables businesses to continue communicating, operating, and innovating.

The talk explores the problems in retail that need text extraction from images with great accuracy.

We explore state-of-art models for text extraction using deep learning and attention networks. We also describe how we have used spatial transformation networks to improve detection of curved text.

The talk ends with a description of how the model and the necessary ML engineering pipelines have been productionized at Walmart scale.

- Principles to build explainable, interpretable, and trustable models

- Parallels between physics-based models and deep learning models

- From conservation laws to invariant models

- Causality as a special case of invariance

- Constraining causal models using domain knowledge

- Application spotlights: anomaly detection and insurance profitability

Companies are frequently faced with large amounts of unstructured text data, like forum comments or product reviews. Important trends can emerge in these datasets, but it can be time-consuming to read through comments, and keyword matching frequently misses critical nuances.

We'll discuss how we've approached this problem at Google using Natural Language Processing, with examples of the approach applied to open datasets.

We'll explore how this fits into the ML project lifecycle, with examples of common pitfalls. Finally, we'll highlight how to use this technology as part of a "human in the loop" approach to supercharge your existing team members.

One of the most important problems in transfer learning is the task of domain adaptation, where the goal is to apply an algorithm trained in one or more source domains to a different (but related) target domain. In many cases, the differences between the source and target domains can be modeled and characterized by a set of variables that are observed under different contexts.

These context variables can often be linked to real-world environmental variables, such as the organism in a genetics study or the hardware platform where a machine learning system is deployed. In this talk, I will present an algorithm based on Markov blanket recovery that enables us to identify invariant causal features across source and target domains.

I will show that our approach scales well to a large number of variables, increases robustness, and makes the task of domain adaptation computationally tractable by exploiting locality in the causal structures. I will also demonstrate some empirical evidence showing the effectiveness of our proposed approach in different domain adaptation scenarios.

• Automating previously manual data processing tasks with ML
• Enabling the measurement of multiple parameter permutations for data processing, integration or feature selection.

• Model Development Lifecycle Basics
• Design Constraints for building Production AI Systems
• How to Navigate these and build the correct solution for you challenge

Training data is the fuel that powers the AI algorithmic engine. Without well prepared data, machine learning cannot take place. Refining that fuel at scale requires industrial-sized operations. Sourcing the data needed for the engine can often be just as complex.

As a result, training data often becomes an operational bottleneck in deploying AI solutions at scale.The process to acquire, label, and verify training data is presently slow, manual and therefore expensive.The biggest challenge today to wider adoption of scalable AI technologies is not better algorithms and models, but instead obtaining more high-quality training data.

AI Models will continue to get more complex and eventually data-starved without the deployment of new paradigms on how to approach a training data problem that today is limited by linear scalability of manual labor. Breaking the brute-force cycle of more data requiring more human annotators in turn producing better training data is a lynchpin to creating cost-effective AI Solutions. A cost-efficient process should combine AI, skilled human annotators, and optimized tools with interaction models that are designed to minimize errors and optimize throughput. Robust QA methodologies throughout the data “assembly line” complete the industrialization approach.

The success or failure of data training, testing, and tuning is realized in the final results of the model. Obtaining better training data is based on at least ten criteria that each need to be addressed and optimized for the task to be automated and/or improved:

• Accuracy and Consistency
• Quality
• Flexibility
• Speed
• Price
• Service and Maintenance
• Security and Privacy
• Reliability and Trustworthiness (No crowd sourcing)
• Scalability
• Ethical Operations

This presentation will highlight, through real examples, how combining the use of AI to automated or semi-automated the annotation process paired with human expertise can intelligently and optimally combine to best satisfy the requirements above. The result will satisfy customer needs and feed AI algorithms with larger and larger volumes of high-quality training data. Methods to be examined here include:

• Optimally combining human expertise and machine learning technologies, depending on their relative strength, the dataset, and the subject domain to increase throughput and data consistency.
• Using technology and a highly customizable platform to quickly adapt data annotation tools and services to a given task.
• Continuously improving data quality and speed as more data is processed.
• Understanding and accommodating the benefits and limitations for a particular company along the spectrum of privacy and security.
• Developing a flexible approach to provide personalized data annotation services and manage data sets across the continuum of customers and automated annotation.

We're at a tipping point of a paradigm shift in the way we will interact with technology. Embodied is aiming to lead this charge through an advanced social interface that respects humans’ natural modes of interaction, beyond simple verbal commands, to enable the next generation of computing, and to power a new class of machines that will change the world around us. Paolo will discuss how he and his team at Embodied are rethinking and reinventing how human-machine interaction is done - starting with the recent announcement of Moxie.

Moxie is an animate companion that helps children build social, emotional and cognitive skills through everyday play-based learning and engaging content developed in association with experts in child development and education. Embodied has assefmbled a world class team of experts in engineering, technology, game design, and entertainment to bring to life a robot with machine learning technology that allows it to perceive, process and respond to natural conversation, eye contact, facial expressions and other behavior as well as recognize and recall people, places, and things.

In this presentation, Paolo Pirjanian will discuss his journey developing Moxie and his vision for how technology can improve and enhance our lives. He will also demo some of the key features in the technology that enable the creation of an effective social interface with fluid conversation, body language, and more.

• Distributing learning processes to build automated AI systems
• Ensembling created models to accelerate next-level training processes
• Architecting communication to facilitate human-AI interactions
• Designing operational structure of agents and humans to arrange their communications
• Evaluating component-level performance of Human-AI teaming to optimize human and AI efficiencies

Healthcare data exists in a variety of formats. These include radiology/pathology images, Clinical reports, Voice Transcripts and patient records in an electronic medical record (EMR) system. Working across these modalities and retrieving a single lens view of a patient from these data formats is challenging. With the advent of cloud computing and the advancements in deep learning, users can now derive insights from these datasets and use this information in the patient’s care profile. In this session, we will explore techniques that allows users to utilize medical images, voice transcripts and clinical notes in machine learning pipelines for optimizing clinical workflows. We will also go over some managed services from Amazon Web Services that allows users to build and scale these pipelines.

We begin by presenting an overview of the data infrastructure at Carbon Health and the Data Science team roadmap. We then walk through our process for publishing the first ever open clinical data repository. We present some of the research efforts that this data enables, including swab type comparisons and disease symptom assessment. We wrap up the first part of the talk by taking a tour of our clinical decision support system with a summary of its implementation.

In the second part of the talk, we take a detailed look at our efforts in risk assessment for helping businesses bring employees back into the office. We ran simulations to test how different parameters would affect outbreaks in the office. We found that generally tests with faster turnaround times (even at the expense of lower sensitivity rates), more frequent testing cadence, and remaining home when displaying symptoms diminished transmission rate.

• Augmenting your existing software application architecture to seamlessly extract electronic patient data from hospital systems, m-Health devices, and electronic medical records, whilst remaining compliant with HIPAA & HL7.
• Ensuring that the data collected is organised, sorted and processed in a way that takes into account the variety of data being collected, and consequently generating actionable insights
• Security best practises in the IT department and beyond – ensuring your critical healthcare data secured from an operational best practise perspective, as well and IT operations perspective.

Making the CXOs sign-up for AI- Evangelizing the idea of AI with the C-suite in your org; getting their buy-in/funding for an aspirational AI based project and running a POC,

Project inception to implementation journey- How to decide whether to do in house or hire a vendor? How to identify the right team? How to prioritize the use cases? How to define the KPIs for an outcome driven project etc

Herbalife Nutrition is a health and nutrition company that works on a multi level marketing business model that involves distributors selling the products to the customers often by word of mouth and direct sales. In such an industry that heavily relies on human connect, how can AI help grow the business? We decided to find ways for AI to help our business drivers i.e. our distributors a) to expand their ability of selling by amplifying their cognitive strengths b) interact with customers to free them for higher-level tasks c) and embody human skills to extend their physical capabilities

Key Learnings - Learnings from an AI project vs a non-AI project. Defying the common myths

Overall, the audience will learn and see journey of executing an AI project end to end all the way from inception to execution through a real example of human + machine partnership. While the example is for Healthcare/Nutrition/Multi-level marketing industry, the learnings apply to pretty much all industries

A Discussion on The Likelihood of an Industry Wide Best Practice for Explainable AI for and Ethical Use of Patient Data

• Suitability and potential to leverage federated learning to train models on-device, on locally available data
• Ethical use of Patient Data – what does responsible use look like?
• Implementing algorithmic assessment into compliance process
• Fine tuning model auditing best practice
• Building out a body of behavioural data from m-health organisations – delivering a patient-centric AI partnership paflan for your organisation

• Generating mass stockpile of performance data from multiple environments and data sources and deploying applications to constantly monitor and feedback where necessary
• Generating real-time visibility of your entire IT and Application architecture to allow or IT teams to respond accurately – and quickly – at the first sign of a genuine threat
• Eliminating the “magic pill” narrative in the market and incrementally building a technical solution to pave the road for your AIops journey

• Overcoming market discontinuity to invest in products, applications and services which integrate with one another
• When is the right time to make an acquisition to augment your AIOps vision? Is the market mature enough yet?
• Which use cases are more proven? Where can the largest ROI be found? How do you derisk your investment as best you can?

Artificial Intelligence workloads process massive amounts of data from structured and from unstructured sources. High performance, architecturally optimized solutions that can harmoniously exist within existing data centers is just one of the infrastructure factors IT teams need to easily address without increasing complexity.

In this talk, we will discuss how to advance your organization’s AI endeavors with simple, scalable, and powerful infrastructure. Maximize data science productivity with easy to use infrastructure that speeds AI development from prototyping to production and learn tips from the experts on how to accelerate your entire data pipeline and power radical AI-fueled business transformation.

We live in a world where applications run on heterogeneous platforms/ technologies, demanding instant scale ups/downs while increasingly beset by security breaches. The traditional approaches to prevent have failed abysmally. We need to fundamentally rethink operations and invert the man-machine paradigm.

Artificial Intelligence and Machine Learning algorithms along with probabilistic matching, fuzzy logic, linguistic correlation and pattern mining can help organizations not only prevent but also anticipate and proactively be ahead of the remediation curve.

We will look at how to build an environment dependency model based on environment topology, component dependencies and configuration dependencies.

This session will additionally look at cutting edge techniques and design and application of algorithms that will help build a proactive monitoring and performance management and remediation platform for multi-cloud management and other physical/virtual resources leading to AIOps.

The agility, iterative approach, and reactive nature of DevOps took the place of the endless preparatory and super conservative planning for even the smallest change typically common in older IT environments. However, the extremely dynamic nature of modern application stacks and deployments has created a new set of problems.

• Register for this presentation as we dive deeper into the:
• Limitations of DevOps for modern systems from the lack of visibility in elastic infrastructures to the difficulty in finding the root cause of service disruption
• Capabilities of Artificial Intelligence to uncover new patterns in operations undetected by traditional tools
• Value of AIOps to identify the root cause and deliver actionable automation for modern applications

• Identifying companies that can easily integrate with others’ / your own IT stack
• Assessing internal talent and identifying strengths and weaknesses
• Is it fair, or even sensible, to give a large amount of consideration to low liquidity in companies with high-growth potential and scalable business models?