Talk Abstract

This talk will provide a broad overview of Google DeepMind's research in audio, image, 3D and video generation. We will focus on autoregressive approaches and diffusion-based approaches. I will try to also provide a historical lens on attempts to use ML to genereate media, going back to early work using RNNs and LSTM in the late 1990s and early 2000s. Given the breadth, we won't have time to dive into many of the mathematical details of these methods.

Speaker Bio

Doug is leading Google DeepMind’s research efforts in Generative Media, including image, video, 3D, music and audio generation. He also leads a broader group active in areas including including Fundamental Learning Algorithms, Natural Language Processing, Multimodal Learning, Reinforcement Learning, Computer Vision and Generative Models. His own research lies at the intersection of machine learning and human-computer interaction (HCI). Doug created Magenta, an ongoing research project exploring the role of AI in art and music creation. He is also an advocate for PAIR, a multidisciplinary team that explores the human side of AI through fundamental research, building tools, creating design frameworks, and working with diverse communities. In the past, Doug worked on music perception, aspects of music performance, machine learning for large audio datasets and music recommendation. He completed his PhD in Computer Science and Cognitive Science at Indiana University in 2000 and went on to a postdoctoral fellowship with Juergen Schmidhuber at IDSIA in Lugano Switzerland. Before joining Google in 2010, Doug was faculty in Computer Science in the University of Montreal machine learning group (now the MILA machine learning lab) where he became Associate Professor. For more information see http://g.co/research/douglaseck.

Talk Abstract

While ML model development is a challenging process, the management of these models becomes even more complex once they're in production. Shifting data distributions, upstream pipeline failures, segmentation challenges, model hallucination and much more can create thorny feedback loops between development and production. In this talk, Niko will: (1) examine naive ML workflows that don’t take the development-production feedback loop into account and explore why they break down; (2) share industry case studies where teams have applied these principles to their production ML systems; (3) showcase generalizable system design principles that help manage these feedback loops more effectively.

Speaker Bio

Niko Laskaris is a researcher and entrepreneur. Previously, Niko worked as a research scientist at Yale University, studying climate change and complex systems. He also founded and built an educational consulting company based in Vermont, Long Trail Prep. Niko holds a bachelor’s degree from Yale University in Connecticut.

Douglas Blank is the Head of Research at Comet ML, where he works with many teams, including Engineering, Customer Success, and Product Design. Prior to Comet, Douglas completed his PhD in Computer Science and Cognitive Science from Indiana University, Bloomington. His thesis explored the training of neural networks to make analogies. He taught courses in Robotics, Cognitive Science, and Computer Science at Bryn Mawr College, where he created a research agenda called "Developmental Robotics," focusing on using Deep Learning as the foundation for a mentally developing robot.