The future of AI lies in collaboration between technologies

Advancements are made continuously in several different AI specialisations, e.g. machine learning, computer vision, robotics, natural language processing for text and speech (NLP), planning, optimisation, knowledge representation, and automated reasoning.

Many find it fascinating that computer vision at present surpasses the ability of doctors to assess X-rays, that humanoid and animal-inspired robots can do somersaults and have better balance than most people, and that computers can compose poetry or distinguish dogs from cats with 98 percent accuracy.

What won’t be possible in ten years?

– “It’s clear that advancements in every field will continue, but you have to remember that there is a difference between doing things in a lab environment and dealing with problems in reality,” says Holst.

– “The fact that AI can differentiate between cats and dogs through image recognition with 98 percent accuracy doesn’t help the processing and manufacturing industry much. An error of two percent can be catastrophic in a manufacturing process. It’s one thing to solve a trivial problem in the lab, but solving real-world problems that require domain knowledge to ensure the right problem is solved is another matter.”

Different technologies have different strengths

In a dynamic and complex real-world environment – especially when there is a lot at stake and nothing is permitted to wrong, autonomous vehicles being a good example – different AI technologies must interact effectively.

• Data analysis and machine learning are valuable for analysing a situation and creating a uniform overview from available information.
• Planning and optimisation may be needed to find the best action given the identified situation.
• Image analysis is needed if you will deal with images, language analysis if you will handle human language.
• If the domain contains complex rules, such as contracts, to which the AI system needs to relate, automated reasoning may be required

According to Holst, only when different AI technologies are combined can computers become truly ‘intelligent’ and function in complex contexts:

– “Intelligence is not a single ability, but is made up of multiple abilities. If you compare it with the human brain, it is not a homogeneous mass – it is divided into several centres: the visual cortex, language centres, the motor cortex, the sensory cortex, the frontal lobes, and so on, which give us language skills, spatial thinking, abstract thinking, motor skills, social skills. If you had only one of these, you would not be regarded as truly intelligent.”

Challenging to get AI technologies to work together

– “Getting the different parts to work together is not always easy either,” explains Holst. “Knowledge of this ‘AI glue’ that helps different AI technologies to interact will also become more important in the coming years.”

One way of understanding what could serve as a ‘glue’ is by looking at an example: SMHI’s weather forecasts. TV weather presenters may say: “It’ll be sunny in Stockholm tomorrow”, and a preschool teacher heeds this forecast and plans the children's outing for the day. The process in this example has no glue. It only works if what the weather presenter said actually happens.

– “SMHI’s forecast models may have actually indicated a 70 percent probability of clear skies, a 15 percent probability of clouds, and a 15 percent probability of rain. If you can obtain probabilities from the forecast and then take them into account when planning, you can attain a better, more robust plan. In this example, it is the probabilities – and how planning can relate to them – that serve as the glue to facilitate planning that will work in reality.”

Difference at system level

So what will actually happen in AI in ten years? Anders Holst predicts that AI will make a difference at system level in factories, for example.

– “Planning production and maintenance, for instance, will be more optimised. Most things will be able to be done more dynamically, with analyses based on multiple variables. For example, when is the right time to stop production, for how long, and which parts must be replaced? At present, this is done according to a predetermined schedule, but AI will be able to find the right times for production and maintenance based on the expected electricity price and the supply and price of raw materials, and determine which compressors should be replaced, and so on.

– “We will see an increased number of autonomous systems that handle all sorts of tasks in industry and society, based on combinations of many different AI technologies.”