Integrated AI for the future
Despite impressive advancements in AI, we are still a long way away from creating systems that can perform complex activities autonomously in real-world domains. Outside tightly controlled or simulated environments, we are not yet able to build systems that exhibit the flexibility, adaptability, and awareness required to perform those tasks which we would prefer AI systems to take on rather than humans, especially those that are prohibitively dangerous, onerous, or mundane for people in a modern society.
However, the currently dominant paradigm of training systems from scratch on each new problem – with data and compute requirements increasing dramatically as we attempt to tackle ever more complex problems – is becoming unsustainable. Without combining and reusing components flexibly “on the fly”, there is a real danger that we will soon hit a hard barrier both in terms of the capabilities of AI systems and their sustainability.
We must now focus on making AI expandable within a context of autonomy that combines sensing, reasoning, and acting in order to overcome this roadblock, and to unlock the potential of developing truly transformative applications in key domains.
Achieving the levels of autonomy required to deploy AI systems safely in everyday life domains requires endowing them with a range of advanced capabilities from context awareness and knowledge reuse to flexible behaviour composition and social ability. Many of these component capabilities have been successfully studied in sub-fields of AI such as:
- machine learning
- natural language processing
- machine vision
- robotics
- symbolic AI
- agent technology.
Yet, recent progress in these areas has come at a price of focusing on single capabilities – a concerted effort is now needed to focus on integrating them.
A major recent success in this direction has been the award of a £3.9 million Turing AI World-Leading Researcher Fellowship to Professor Mirella Lapata from our School of Informatics to create the Edinburgh Laboratory for Integrated Artificial Intelligence, which will focus on developing new types of AI systems that focus on integrating these capabilities, as well as teaching machines new skills such as how to make generalisations, deal with changing situations, and be creative, by writing poems, for example.
Big questions we want to focus on in the long term in this direction include:
- How should AI systems select appropriate data sources to focus their attention on?
- How can we enhance them using human knowledge?
- How can they exploit previous solutions to solve new problems?
- How can they assess the quality of their solutions, knowing when to seek human guidance?
- How should we decide whether and to what extent we should allow AI systems to act autonomously?
We believe that a sustained focus on these challenges has the potential to deliver important breakthroughs in terms of our ability to build systems that can learn fast, utilise prior knowledge, interpret information, and combine skills much more competently than they are presently able to; and will be fundamental to shaping the “next wave” of AI technologies.