Invited Talks
Timnit Gebru

Abstract: On December 2nd, I was fired from Google citing an email I wrote regarding the company’s treatment of women and Black people. Leading up to my firing I had been asked to retract a paper titled On the Dangers of Stochastic Parrots: Can Language Models be Too Big?, and after my firing public communication from Google Research SVP Jeff Dean claimed that this work “didn’t meet our bar for publication.” What transpired afterwards was a barrage of harassment, stalking and slurs hurled at me, my collaborators and members of the Ethical AI team, with my former co-lead Margaret Mitchell being fired on February 19. In this keynote, I will go through some of the key points in the Stochastic Parrots paper, highlighting what happened to me and my collaborators as examples of these key points. My firing follows other high profile firings of Black women such as Dr. Aysha Khoury from Kaiser Permanente School of Medicine, recruiter April Curley from Google, and most recently lecturer Khadijah Johnson from Cornell Tech for speaking up against injustice towards our communities. This is happening as machine learning based models are amplifying negative medical outcomes for Black people, and social media …

Yejin Choi

Despite considerable advances in deep learning, AI remains to be narrow and brittle. One fundamental limitation is its lack of commonsense intelligence: trivial for humans, but mysteriously hard for machines. In this talk, I'll discuss the myth and truth about commonsense AI---the blend between symbolic and neural knowledge, the continuum between knowledge and reasoning, and the interplay between reasoning and language generation

Michael Bronstein

For nearly two millennia, the word "geometry" was synonymous with Euclidean geometry, as no other types of geometry existed. Euclid's monopoly came to an end in the 19th century when multiple examples of non-Euclidean geometries were constructed. However, these studies quickly diverged into disparate fields, with mathematicians debating the relations between different geometries and what defines one. A way out of this pickle was shown by Felix Klein in his Erlangen Programme, which proposed approaching geometry as the study of invariants or symmetries using the language of group theory. In the 20th century, these ideas have been fundamental in developing modern physics, culminating in the Standard Model.

The current state of deep learning somewhat resembles the situation in the field of geometry in the 19h century: On the one hand, in the past decade deep learning has brought a revolution in data science and made possible many tasks previously thought to be beyond reach -- including computer vision, playing Go, or protein folding. At the same time, we have a zoo of neural network architectures for various kinds of data, but few unifying principles. As in times past, it is difficult to understand the relations between different methods, inevitably resulting …

Manuela Veloso

AI enables principled representation of knowledge, complex strategy optimization, learning from data, and support to human decision making. I will present examples and discuss the scope of AI in our research in the finance domain.

Lourdes Agapito

As humans we take the ability to perceive the dynamic world around us in three dimensions for granted. From an early age we can grasp an object by adapting our fingers to its 3D shape; understand our mother’s feelings by interpreting her facial expressions; or effortlessly navigate through a busy street. All these tasks require some internal 3D representation of shape, deformations, and motion. Building algorithms that can emulate this level of human 3D perception, using as input single images or video sequences taken with a consumer camera, has proved to be an extremely hard task. Machine learning solutions have faced the challenge of the scarcity of 3D annotations, encouraging important advances in weak and self-supervision. In this talk I will describe progress from early optimization-based solutions that captured sequence-specific 3D models with primitive representations of deformation, towards recent and more powerful 3D-aware neural representations that can learn the variation of shapes and textures across a category and be trained from 2D image supervision only. There has been very successful recent commercial uptake of this technology and I will show exciting applications to AI-driven video synthesis.

Kate Saenko

Is my dataset biased? The answer is likely, yes. In machine learning, “dataset bias” happens when the training data is not representative of future test data. Finite datasets cannot include all variations possible in the real world, so every machine learning dataset is biased in some way. Yet, machine learning progress is traditionally measured by testing on in-distribution data. This obscures the real danger that models will fail on new domains. For example, a pedestrian detector trained on pictures of people in the sidewalk could fail on jaywalkers. A medical classifier could fail on data from a new sensor or hospital. The good news is, we can fight dataset bias with techniques from domain adaptation, semi-supervised learning and generative modeling. I will describe the evolution of efforts to improve domain transfer, their successes and failures, and a vision for the future.

Kyu Jin Cho

Robotic technologies are coming closer into our everyday life. Unlike conventional robots with metal body and rigid joints, robots used in everyday life must have a form factor that considers usability as well as functionality. Designing a robot with soft body will enable robots to be more human-friendly, but there are several challenges. I will share some general thoughts and strategies for designing soft robots and how embodied intelligence can enable simple yet intelligent design of robots to be used in everyday life.

Alyosha Efros

Why do self-supervised learning? A common answer is: "because data labeling is expensive." In this talk, I will argue that there are other, perhaps more fundamental reasons for working on self-supervision. First, it should allow us to get away from the tyranny of top-down semantic categorization and force meaningful associations to emerge naturally from the raw sensor data in a bottom-up fashion. Second, it should allow us to ditch fixed datasets and enable continuous, online learning, which is a much more natural setting for real-world agents. Third, and most intriguingly, there is hope that it might be possible to force a self-supervised task curriculum to emerge from first principles, even in the absence of a pre-defined downstream task or goal, similar to evolution. In this talk, I will touch upon these themes to argue that, far from running its course, research in self-supervised learning is only just beginning.