Three broad technological trends have driven the application of AI in the Sciences: improved computer hardware and processors; improved data availability across the sciences; and improved AI software and new machine learning. The growing convergence of these accelerating trends provides a once in a generation opportunity to accelerate scientific discovery for the public good, and to create synergies across scientific areas.
However, the application of AI in scientific discovery presents us with very different challenges compared to games such as chess, shogi, and Go. Kitano (2002) argues that firstly, scientific discoveries require an order of magnitude larger hypothesis space, and a more elaborate and expensive verification process. Secondly, there are still challenges in the integration of devices and modules, and the development of integrated AI systems that can execute experiments autonomously (King et al., 2008). Third, while many possible discoveries could be made by an AI scientist, we need to ensure that they are relevant to human civilization and aligned with our values. This capability requires a broader knowledge about the world, a deep understanding of the human scientific discovery process and good reasoning about potential applications and social implications.
We envision future AI systems that do not have these limitations and will, therefore, have the potential to transform scientific discovery and enable an unprecedented expansion of knowledge. Building on the legacy of the Turing Test and the RoboCup model, we invite the community to join us in the Turing AI Scientist Grand Challenge: developing AI systems capable of making Nobel quality scientific discoveries highly autonomously at a level comparable, and possibly superior, to the best human scientists by 2050.
The clear and explicit goal of the challenge is to develop an AI system that has highly autonomous capabilities to perform research, rather than developing AI tools. Human scientists already use a wide range of AI tools for specific tasks that contribute to discoveries (e.g., data analysis, text extraction, etc.). However, these AI tools cannot be considered AI Scientists because they only address narrow aspects of the scientific process and they lack the autonomy of human scientists in setting their goals, interpreting results, and communicating findings.
Along the way, intermediate grand challenges will need to be defined to structure specific problems and demonstrate tangible progress. A key criterion for these challenges will be the synergistic aspects of discovery in scientific domains, and the development of the AI Scientist technology. Achieving the Turing AI Scientist Grand Challenge will take several decades of hard research, but we expect that it will drive progress in both science and AI research.
Explaining the science
The challenge requires an integrated AI system with very diverse capabilities. The AI system must be able to make a strategic choice about research goals, design protocols and experiments to collect the necessary data, notice and characterise a significant discovery in terms that other scientists will comprehend, communicate in the form of publications and other scientific means to explain the innovation and methods behind the discovery and articulate the significance of the discovery in terms of its novelty with respect to state of the art and of its impact in terms of societal benefits.
We envision intelligent systems that can support scientists in the areas of (Gil et al. 2018):
Knowledge Representation and Capture: Developing expressive and efficient frameworks to represent and capture a wide range of scientific knowledge about processes, models, and hypotheses. Capturing scientific knowledge will push the limits of state of the art. This work is expected to have a significant impact on the reproducibility of science, as automating science requires semantic precision.
Sensing and Robotics: Prioritising data collection based on the scientific knowledge available. The trend in science is for more and more of the physical execution of experiments to be done using robotics. This will increase the productivity of science as robots work cheaper, faster, more accurately, and longer than humans.
Information Integration: Representing data and models as a “system of systems” where all knowledge is interconnected.
Machine Learning: Enriching algorithms with knowledge and models of the relevant underlying processes. Methods will need to be enriched with models of the relevant science, and their results explainable to scientists. How best to include background knowledge in machine learning is one of the most important open questions in the field.
Interfaces and Interactive Systems: Exploring and understanding user context using interconnected knowledge. The AI systems’ interactions with scientists must be guided by a knowledge-rich user model that enables the AI systems to act as colleagues. Information from human scientists must be able to be understood by the AI systems, and the AI systems must be able to explain their conclusions to human scientists.
Starting in January 2021, we will:
Create a landscaping study for AI systems that have highly autonomous capabilities to perform scientific research. This includes defining challenges for AI research that will push discoveries beyond the state of the art and scope collaborative research work for AI and experts from other disciplines.
Generate a multi-year roadmap with technical quests and intermediate milestones for scientific discovery in biomedicine, environmental sciences, and materials science.
Facilitate coordination, resource development and sharing.
Create distributed collaborative AI hubs around the world (US-UK-Japan).