By combining extraordinary intellectual freedom and scientific rigour with access to top resources and a structured, supportive culture, we have established an unparalleled track record of AI breakthroughs.
Our pioneering scientists and engineers have taught agents to cooperate, play world-class chess, diagnose eye disease, and predict the complex 3D shapes of proteins. Combined with a strong focus on safety, ethics, and robustness, the team works to create systems that can provide extraordinary benefits to society.
General purpose learning systems must be able to cope with the richness and complexity of the real world. These topics drive the control and robotics teams at DeepMind, which aim to create mechanical systems that can learn how to perform complex manipulation tasks with minimal prior knowledge. The shared ambition is to create systems that are data-efficient, reliable, and robust.
The development and use of deep neural networks underpins much of the current wave of AI research and is a critical technique for many modern applications such as machine translation. Deep learning methods are at the core of many research areas at DeepMind, including deep reinforcement learning, generative models, theory and optimisation, transfer learning, computer vision, program synthesis, and hierarchical reinforcement learning.
The brain is the best example of a general purpose learning system and we use it as an inspiration for our algorithms. We conduct experiments to try and understand how human intelligence works, from memory and learning to internal navigation systems and motor control. Their insights are then used to build the next generation of algorithms. We also develop tools inspired by neuroscience that can probe our AI systems in the same way a neuroscientist studies neural circuits in the brain - an important step towards building interpretable AI systems.
Giving computer systems the ability to learn through trial-and-error has shaped many of DeepMind’s most well-known projects including AlphaGo, AlphaZero, and AlphaStar. We continuously push the boundaries of this powerful technique, advancing areas such as credit assignment, planning, locomotion, and meta-learning.
We study theoretical and practical problems that might arise when building general purpose learning systems. These problems fall loosely into three categories: specification (defining the purpose of a system), robustness (designing systems that can withstand outside perturbations), and assurance (monitoring and controlling a system’s activity). Our goal is to understand the behaviour of systems, including unintended behaviours or side effects; aligning agents with the goals, preferences, and ethics of the system's operators; understanding the ways in which artificial intelligence might want to modify itself over time; and approaches to containing or restricting the scope, behaviour, or design of a system.
We focus on the theoretical foundations of machine learning to understand the limits of current architectures and support the development of new, efficient, and effective learning algorithms. Our researchers cover a wide range of topics including passive, active, partial, and full information feedback learning, as well as representation, supervised, and unsupervised learning. In all cases, we aim to create principled solutions that are robust and scalable.
Unsupervised learning is a powerful technique that allows systems to learn directly from datasets that don’t have specific labels or rewards. This is an important attribute for AI, allowing them to learn and therefore make sense of their environment in much the same way a child learns through play and observation. We work on various approaches to generative and predictive models of unstructured data streams, such as text, image and video.
Raia worked in philosophy and religion before switching to machine learning and robotics. Her PhD from NYU focused on representation learning and robot navigation, using convolutional networks to see the world.
Raia’s team researches embodied and lifelong learning in complex situations, including dexterous manipulation with multi-sensor robot hands, robot locomotion, and city-scale navigation.
Ali holds a PhD in generative models from the University of Edinburgh, conducted his postdoc at Microsoft Research in Cambridge, and was a visiting researcher at the University of Oxford.
Ali figures out how computers can learn to see with less supervision. His work involves a mix of deep learning, probabilistic inference, and reinforcement learning.
Jess holds a PhD in psychology from the University of California, Berkeley, and earned her BS and MEng in computer science from MIT.
Jess currently applies insights from cognitive science to problems in AI, with an emphasis on structured representations, model-based reasoning, and planning.
Rich researched reinforcement learning at universities in Alberta and Massachusetts, and at corporate labs within AT&T and GTE, since 1978.
Rich works between DeepMind Alberta and the University of Alberta. He identifies unknown parts of the mind, which therefore prevent us from recreating its abilities in machines.
Remi worked at Inria and taught at École Polytechnique. He did his postdoc at Carnegie Mellon University and holds a PhD on reinforcement learning.
Remi focuses on deep reinforcement learning and combinations with unsupervised and imitation learning, and learning from a teacher.
Yazhe studied theoretical and applied mechanics and civil engineering. She started her career as a civil engineer, but soon found her passion in computer science and machine learning.
Yazhe collaborates with research scientists on advancing our understanding of machine learning and developing state-of-the-art deep learning algorithms.
Hado studied cognitive artificial intelligence and holds a PhD in AI from Utrecht University, NL. He later joined DeepMind after working with Professor Rich Sutton at the University of Alberta.
Hado builds systems and solves challenges with reinforcement learning, deep learning, and optimisation. He also co-leads an effort on core reinforcement learning algorithms.
Jonathan earned a masters in machine learning from the University of Edinburgh, working on modelling sequential data, robot navigation, and climate research.
Jonathan collaborates with his team on cutting-edge machine learning problems, running experiments, discussing new ideas on the whiteboard, or presenting his latest work.
Edward worked in quantitative finance for twenty years, developing skills in mathematics, statistics, and software engineering, which he applies to cutting-edge AI research.
Edward helps organise the research team and contributes to research efforts. He runs the Research Engineering Intern programme and develops agents that learn to collaborate.