As AI continues its infiltration into our daily lives, at what point do we ask the questions “Can and should we trust AI?” From overt AI misinformation such as deepfake videos, to more subtle decision manipulation by biased data and algorithms, what guardrails must we put in place before we can trust the AI behind the systems running our businesses and the apps that run our lives?

Human trust in technology is based on our understanding of how it works and our assessment of its safety and reliability. To trust a decision made by an algorithm, we need to know that it is reliable and fair, that it can be accounted for, and that it will cause no harm.

This talk explores diverse approaches to achieve fairness, robustness, explainability, accountability, value alignment, and how to integrate them throughout the entire lifecycle of an AI application.

If we can agree that trust is essential for AI use and adoption, then what organizational characteristics and best practices should we adopt to foster the creation of trustworthy AI? In this talk we’ll examine pain points, motivators, and dos and don’ts for delivering fair, accountable, and transparent AI based on interviews with Responsible AI industry practitioners.

AI technology is not constrained by the physical borders that divide countries and continents. As AI technologies are being developed world-wide, how will we ensure a baseline of accountability and trust in our systems universally across the globe? Are universal guidelines a realistic possibility, or will AI systems be brought down to the lowest common denominator of trust?

Explore in greater depth the issue of trust with our panel of experts. The panel will respond in real-time to the questions on the minds of our attendees and MIT Technology Review.

Deepfake videos have become increasingly sophisticated, accessible, and- most importantly- convincing, while simultaneously surfacing a set of challenging policy, technology, and legal issues. We begin this segment with an analysis of the state of the art of AI deception.

Deepfakes, AI-generated images, have captured the public imagination with examples of image and video manipulation that was, until recently, the purview of artists and studios. Deepfakes along with their audio analog -- audiofakes -- have the potential to create an "information disorder" where there can be an inability to show something real as real and an ability to show something fake as real. In this talk, we'll contextualize multimedia fakes in the broader misinformation landscape and explore some state-of-the-art solutions for detecting deepfakes and audiofakes at AI Foundation and beyond.

A necessary prerequisite to detecting deepfakes is the ability to quickly and autonomously process video assets. However, training a video recognition model can take up to 50 times more data and 8 times more processing power than training a static image classification model. This talk explores a ground-breaking new technology that will significantly increase video processing speeds, improving our ability to deal with the ever increasing volume of deepfake videos.

Deepfake technology exists and is here to stay. Researchers and technology creators can no longer hide behind the simplistic perspective that “there is nothing I can do if people misuse my work.” This talk gives a short tour of the practical steps that can be taken to limit the harm of synthetic media tools.

Explore in greater depth all aspects of deepfake prevention. The panel will respond in real-time to the questions on the minds of our attendees and MIT Technology Review.

From toothbrushes to toasters, engineers are exploring ways to integrate artificial intelligence into the tools we use every day. Many of these advancements have had great positive impact. But do we really need AI everywhere? And are there unintended consequences of handing over more and more decision-making power to machines?

One of the beachheads of AI adoption has been in the realm of customer service. Chatbots and similar AI interaction technologies are taking over the front line of customer service. What’s the next step in this evolution? This talk examines how the AI research fields of neuroscience, cognitive science, and developmental psychology are combining to take the next step in personalized, AI-driven, client interaction.

Natural language understanding (NLU) is one of the hottest research topics in artificial intelligence, aiming to explore how computers understand human language to complete various text- and semantics-related tasks. In recent years, pre-training language models from Google and Baidu have achieved exciting breakthroughs in NLU. In this session, we will explore some basic challenges with NLU, present Baidu’s latest research projects to address these challenges, and share examples of the implementation of NLU in real-world scenarios.

When an X-ray is taken on a patient, especially in a triage situation, the time that it takes to analyze and interpret the image is a crucial, where minutes and hours matter. Radiologists often have to review hundreds of exams per day, resulting in delayed diagnosis of critical conditions such as a pneumothorax, or collapsed lung. This talk introduces the first FDA-cleared mobile device for x-ray triage and examines the implications for the future of AI-assisted healthcare.

There are so many fascinating examples of AI innovation. Before our scientists dive deep into their work, let’s take a look at the broader landscape of cutting-edge AI and ponder for a moment about what possibilities lie ahead.

Deep learning models perform poorly on tasks that require commonsense reasoning. The underlying cause is that reasoning often necessitates some form of world-knowledge or reasoning over information not immediately present in the input. How can we collect human explanations for commonsense reasoning to improve AI decision making? This talk focuses on language-related use cases for customer service, search, question answer, self-help, and consumer finance.

ML pioneer Andrew Ng has called transfer learning “the next driver of ML commercial success.” Transfer learning makes powerful systems more reusable and reduces the amount of training data, compute, and professional services needed. Is it ready for business deployment or is it still emerging technology? How is it used in business today?

The use of robotics has flourished in industrial settings for many years, due in part to the controlled and known environment of workshop floors. But what will it take for the next generation of robotics to enable them to thrive in messy environments such as a kitchen? This talk explores how advances in deep learning, perception, and sensing are powering the next generation of robotics.

We’ll spend the final minutes of the day with the MIT Technology Review editorial team to get their post-game analysis on everything discussed on stage. Find out what they learned, what surprised them, and what needs further explanation in this end of day recap.

We have a packed agenda planned for the day. Let’s take a look at today’s roadmap and then get right to it.

Since its earliest days, artificial intelligence has been long on promise, short on delivery. Getting to artificial general intelligence isn’t some small step from where we are now; instead it will require an immense amount of foundational progress, something entirely different. This talk is about why AI, so far, hasn’t been on the right track, and what we might do to work toward AI that is capable of functioning in a complex and ever-changing world so that we can genuinely trust it with our homes, our parents and our children, our medical decisions, and our lives.

Moving from narrow AI to more general AI is a reasonable next step that requires reason. Current deep learning AI-based models are excellent at identifying objects without understanding a thing about them. Symbolic AI gives us a way to capture common sense reasoning and domain knowledge about those objects. This talk examines symbolic neural learning AI systems, which may be the next natural step on the path towards artificial general intelligence.

Simulation allows organizations to create virtual environments that replicate real world conditions, enabling them to test their AI algorithms. This highly scalable approach enables testing of multiple complex scenarios in parallel, incorporating any edge cases, safely and securely. This talk explores the role simulation can play in the development of robust AI.

Explore in greater depth the complex issues around robust AI and the road to AGI. The panel will respond in real-time to the questions on the minds of our attendees and MIT Technology Review.

Building ethical artificial intelligence is an enormously complex task. It gets further complicated when one realizes that bias is in the eye of the beholder. Is an AI-based college university admission system that balances applicant acceptances based on gender and geography any more or less biased than one that does not? While we can probably agree that a balanced system is better, who has the authority to make these decisions?

How can we create digital products and services that people can—and do—trust? It’s becoming increasingly important as people become more aware of the possible consequences of data being recorded, joined up, and used by organizations. Designing for trust is a complex, multifaceted challenge, and solutions will be context-specific. However, there is a reliable way of figuring out what works in each specific instance, and that’s by taking a design approach to the problem.

How can enterprises trust AI-generated explanations when it is nearly impossible for a human to understand how deep neural networks make decisions? To date, there’s been limited assessment of explainability methods within the nascent deep learning field, and most existing evaluations focus on subjective visual interpretations. This talk explores a machine-centric strategy for quantifying the performance of explainability methods on deep neural networks.

Hiring is broken. The average job receives 250 applications, yet the candidate chosen by the company fails 30-50% of the time. This talk examines a new approach that removes the biases inherent in the hiring process by using neuroscience games and bias-free AI to predictively match people with jobs where they will perform at the highest levels.

Artificial intelligence has given us algorithms capable of recognizing faces, diagnosing disease, and, of course, crushing computer games. But even the smartest algorithms can sometimes behave in unexpected and unwanted ways, for example picking up gender bias from the text or images they are fed. This talk discusses a framework to prevent aberrant behavior in machine learning by specifying guardrails in the code from the outset. It aims to be particularly useful for non-experts deploying AI, an increasingly common issue as the technology moves out of research labs and into the real world.

Explore in greater depth all aspects of AI bias with our panel of experts. The panel will respond in real-time to the questions on the minds of our attendees and MIT Technology Review.

It’s time to get down to business and examine some real-world applications of AI at work. On the job, AI is augmenting intelligence,and providing data analysis and decision making beyond the capabilities of workers alone.  

While examples of AI implementations at digital-first organizations are fairly well known, less attention is paid to AI’s rise outside the software industry. This talk examines AI’s impact on the processes and bottom lines of non-digital organizations and outlines a pragmatic approach to AI adoption across projects ranging from solution-based deployments to enterprise-wide transformations.

The security industry has embraced AI as a key tool in cyber-defense, but while AI can strengthen defenses, it can also enhance the effectiveness of a cyber attacker. The adversarial nature of attack and defense necessitates that security AI be prepared to address the challenge of purposeful deception. This talk will examine this unique dynamic and its resulting impact on AI models for security.

Autonomous vehicles promise to transform personal mobility and potentially reshape our communities. But personal mobility is just one area this technology will impact. Long haul transportation, last mile delivery, even commerce as a whole may be completely altered when vehicles can drive themselves. In this talk we'll explore some of the non-passenger applications of autonomy, in particular last mile logistics.

Examples of AI’s impact on digital business abound. We’ve all had Netflix recommend movies to us, Amazon recommend products, and LinkedIn recommend jobs. But AI’s reach is now far beyond the digital realm, and our final focus of the day is on unconventional uses of AI. Will these applications someday become the new norm? 

The term “AI Chips” has a different meaning at Frito-Lay than at most other organizations. At Frito-Lay, the term refers to the use of AI in the production of potato chips, and more specifically how AI integration into the manufacturing process has saved the organization millions of dollars. This talk explores the use of machine learning both on and off the manufacturing line at PepsiCo and its subsidiary Frito-Lay.

On April 13th, 2019, OpenAI Five became the first AI system to defeat world champions at an e-sports game. The game of Dota 2 presents novel challenges for AI systems - such as long time horizons, imperfect information, and complex, continuous state-action spaces - all challenges which will become increasingly central to more capable AI systems. By defeating Team OG, the Dota 2 world champion, OpenAI Five demonstrates that self-play reinforcement learning can achieve superhuman performance on a difficult task.

MuseNet is a deep neural network that can generate 4-minute musical compositions with multiple instruments, combining styles from country to Mozart. Not explicitly programmed with our understanding of music, MuseNet discovered patterns of harmony, rhythm, and style by learning to predict the next token in hundreds of thousands of digitized music files. For our last session of the day, let’s experience the sound of AI.

We’ll spend the final minutes of the day with the MIT Technology Review editorial team to get their post-conference analysis on everything discussed on stage. Find out what they learned, what surprised them, and what needs further explanation in this end of conference recap.