Confirmed invited speakers
- Ryan Adams (Harvard University)
Designing Molecules with Deep Learning and Bayesian Optimization
Experimental exploration of new materials is extremely expensive and time consuming, and even computational simulation can take days or weeks. In this talk I will describe my group’s ongoing collaboration with Alan Aspuru-Guzik in the Chemistry Department to use machine learning techniques to accelerate the design of novel organic materials. This work uses ML at a variety of levels, from predicting quantum properties without, e.g., density functional theory computation, to the use of Bayesian optimization and experimental design techniques to make choices about new materials to explore. I will describe some of these techniques and show some recent results from the project.
Ryan Adams is an Assistant Professor of Computer Science at Harvard. He received his Ph.D. in Physics at Cambridge as a Gates Scholar. He was a CIFAR Junior Research Fellow at the University of Toronto before joining the faculty at Harvard. He has won paper awards at ICML, AISTATS, and UAI, and his Ph.D. thesis received Honorable Mention for the Savage Award for Theory and Methods from the International Society for Bayesian Analysis. He also received the DARPA Young Faculty Award and the Sloan Fellowship. Ryan is the CEO of Whetlab, a machine learning startup, and co-hosts the popular Talking Machines podcast.
- François Pachet (Sony)
Modeling and Manipulating style: a new challenge for Artificial Intelligence
Style modeling is becoming an area of research in itself in Artificial Intelligence, due to the increasing needs for content generation application. I will present some results obtained in the ERC funded Flow-Machines projects, in which we attempt to capture, model, and exploit "style" at various levels of the musical and text creation process. I will show some results in modeling style with Markov chains, and constraining the generation to satisfy various structural properties, necessary to make sequences look natural. The techniques developed exploit interesting connexions between statistical modeling and combinatorial optimization. Some of these results, sometimes striking, also raise new questions about the nature of "interesting" sequences, and the mystery of creativity.
François Pachet is director of the SONY Computer Science Laboratory Paris, where he leads the music research team. He received his Ph.D. and Habilitation degrees from Université Pierre et Marie Curie (UPMC). He is a Civil Engineer (Ecole des Ponts and Chaussées) and was Assistant Professor in Artificial Intelligence at UPMC until 1997. He joined the Sony Computer Science Laboratory in 1997 and created the music team to conduct research on interactive music listening, composition and performance. Since its creation, the team developed several award winning technologies (constraint-based spatialisation, intelligent music scheduling using metadata) and systems (MusicSpace, PathBuilder, Continuator for interactive music improvisation, etc.). François Pachet has published intensively in artificial intelligence and computer music. He is co-chair of the Ijcai 2015 special track on Artificial Intelligence and Arts, and has been elected Eccai fellow in 2014. His current goal, funded by a ERC Advanced grant, is to build computational representations of style from text and music corpora, that can be exploited for personalized content generation. He is also an accomplished musician (guitar, composition) and has published two music albums (in jazz and pop) as composer and performer.
- Javier González (University of Sheffield)
Bayesian optimization for synthetic gene design
Biopharmaceutical companies often use mammalian cells as 'factories' to manufacture in silico therapeutic proteins, e.g. cancer medicines such as Avastin. A novel approach to optimize this process is the use of synthetic genes, that once `inserted' in the cells are able to scale-up the production of the proteins of interest. In my talk I will describe our joint project with the Department of Chemical and Biological Engineering in which we are addressing the problem of synthetic gene design using Machine Learning techniques. The two key Machine Learning aspects of the project are: the use of Gaussian Processes as a surrogate model for the cell behaviour and the application of Bayesian Optimization ideas to obtain optimal gene design principles. I will discuss the most relevant challenges of this project and show some preliminary results.
Javier Gonzalez is a research associate in the Machine Learning group of the University of Sheffield. He received his PhD in Mathematical Engineering at Carlos III University of Madrid (UC3M) and was a post-doc at the Johann Bernoulli Institute for Mathematics and Computer Science of the University of Groningen. Javier works on non-parametric Machine Learning techniques with a particular interest on problems in Computational Biology. His most recent work focuses on various methodological aspects of Bayesian Optimization with applications to problems in Synthetic Biology and scalable biomanufacturing processes. He is the core developer of GPyOpt, a python framework of Bayesian Optimization that allows for parallel computation and the analysis of complex global optimization problems.
- Michele Sebag (Université Paris-Sud),
The human in the Loop
In some ill-defined or under-specified contexts (e.g. robotic behavior in an open world, interactive optimization, or adaptive interfaces), the learning agent would benefit from having the expert/user in the loop, indicating her preferences about the most appropriate agent behavior. Learning from the user in the loop raises interesting issues. A first challenge (akin active learning or rather Bayesian optimization) is to ask as few questions to the expert as possible, while preserving a good trade-off between exploration and exploitation. A second challenge is to cope with the expert's inconsistency (error is human and the expert's preferences usually evolve along the process). A third challenge is that the human behavior is influenced by the agent behavior...
With a background in maths (Ecole Normale Supérieure), Michèle Sebag went to industry (Thalès) where she started to learn about computer science, project management, and artificial intelligence. She got interested in AI, became consulting engineer, and realized that machine learning was something to be. She was offered the opportunity to start research on machine learning for applications in numerical engineering at Laboratoire de Mécanique des Solides at Ecole Polytechnique. After her PhD at the crossroad of machine learning (LRI, Université Paris-Sud Orsay), data analysis (Ceremade, Université Paris-10 Dauphine) and numerical engineering (LMS, Ecole Polytechnique), she entered CNRS as research fellow (CR1) in 1991. In 2001, she took the lead of the Inference and ML group, now ML & Optimization, at LRI, Université Paris-Sud. In 2003 she founded together with Marc Schoenauer the TAO (ML & Optimization) INRIA project. Her research interests include reinforcement learning, preference learning, information theory for robotics and surrogate optimization.