2 International AI experts: Towards the third wave of artificial intelligence
Where is AI going and what opportunities does the development create for Finland? Nine distinguished international experts were asked to give their answers to these questions. Coming from different backgrounds, they are well placed to analyse extensively the future perspectives of artificial intelligence. This section is based on the interviews with the experts and reflecting their views.
- Christian Guttmann, Vice President, Global Head of Artificial Intelligence & Data Science, Tieto, Executive Director, Nordic Artificial Intelligence Institute, Professor (adj. assoc.), University of New South Wales, Australia. Senior Researcher, Karolinska Institute, Sweden
- Gesche Joost, Professor for Design Research at the Berlin University of the Arts, Head of the Design Research Lab, Germany
- Doina Precup, Research Team Leader, DeepMind and Associate Professor, McGill University, Canada
- Michele Sebag, Professor, Deputy director of Laboratoire de Recherche en Informatique, Head of A O team, CNRS, France
- John Shawe-Taylor, Professor, Head of Department of Computer Science, University College London, UK
- Jim Spohrer, Director, Cognitive OpenTech at IBM, California, USA
- Masashi Sugiyama, Director of RIKEN Center for Advanced Intelligence Project, Japan and Professor for Machine learning and statistical data analysis at the Department of Complexity Science and Engineering, the University of Tokyo,
- Volker Tresp, Professor for Machine learning at the Ludwig Maximilian University of Munich and Distinguished Research Scientist at Siemens, Germany
- Harri Valpola, Founder and CEO, Curious AI, Finland
The interviews were conducted and the summary based on them compiled by:
- Samuel Kaski, Academy Professor at Aalto University, Director of the Finnish Center for Artificial Intelligence (FCAI)
- Heikki Ailisto, Research Professor at VTT Technical Research Centre of Finland
- Arho Suominen, Senior Scientist at VTT Technical Research Centre of Finland
Over the past few years, the biggest successes in AI have been achieved in machine learning and especially in deep neural networks. Huge amounts of data, good development tools and the computing power that is increasing each year have boosted the development. Large companies, such as Facebook, Google, Amazon, Alibaba, Baidu and Tencent make effective use of AI technologies and they are also investing heavily in AI research.
Dividing the development of AI into waves helps us better understand the phenomenon. John Launchbury from DARPA proposes division into three waves:
- handcrafted knowledge;
- statistical learning; and
- contextual adaptation.
The first wave is also called the symbolic or classic artificial intelligence because it was at the core of the AI research from the 1960s to the 1980s. Statistical learning is now the dominant technology and it is taking place by means of machine learning and deep neural networks. The third wave is expected to emerge in the 2020s.
According to the nine experts, machine learning and, in particular, deep neural networks will also be the most important AI technology in the foreseeable future (in the coming years). The most popular methods, which involve what is called supervised machine learning, require a large amount of high-quality training material accompanied by correct answers (labeled data). Using such material, the methods also learn to produce answers to new inputs. Consumer sector actors, such as retail, digital entertainment and social media companies, have access to a large amount of data suitable for training purposes and this partially explains why they are leaders in AI applications.
If you want to teach neural networks to identify a cat in a picture, there is a fairly large number of pictures available in which the word ‘cat’ is connected with the picture as metadata. This is, however, not the case in all applications. For example, the measurement data saved from industrial processes has not always been supplemented with metadata. This is also often the case with healthcare data. A sufficient amount of data may be available but it is not accompanied by classification data. At the same time, producing metadata manually during the measurement process or afterwards requires substantial resources. The data may also be of poor quality or unreliable. For example, in an industrial plant, the sensor may be defective or the sensor placed on the patient’s skin in a hospital may momentarily get detached and give a wrong reading.
In other words, there are many industries that do not have access to high-quality data masses supplemented with metadata. This restricts the use of existing AI technologies. Recent advances in machine learning have provided partial solutions to the problem in the form of transfer learning, reinforcement learning and simulated training data. One and few shot learning as well as weakly supervised learning are also potential solutions. Many experts believe that weak supervision or unsupervised learning may offer partial solutions to the challenges posed by data- based learning.
In five years we may have AI that can do basic commonsense reasoning, and there are leaderboards to track the progress.
– Jim Spohrer
HOW CAN THE PROGRESS BE MEASURED?
We must be able to monitor and measure advances in AI technologies and in their performance. Scientific publications tell us how much is invested in research content, while the investments made by companies provide an indication of the economic inputs into artificial intelligence. Methodological performance can be monitored with tests on standardised databases and leaderboards connected with them. The leaderboards tell how the results are evolving and which of the research teams is at the top.
Experts are predicting that there will be a technological revolution when the existing machine learning methods reach their limits. The pendulum may swing back from data-oriented methods to symbolic methods. The third wave of artificial intelligence will combine unsupervised learning with the methods of symbolic artificial intelligence, such as reasoning, semantic representation, logic and search techniques.
Finland may get a head start in the third wave of artificial intelligence because we have a strong research tradition in unsupervised learning. By strengthening existing top expertise and by investing in domains important to Finland, we can enhance our position in the global AI competition. It should also be noted that funding should also be channelled to information technology and computer science research taking place outside the core of artificial intelligence. This is because we can never be sure where the next breakthrough takes place.
The third wave of AI will combine current statistical and symbolic methods with unsupervised learning.
– Harri Valpola
One interesting aspect is the level of automation and independence of the AI systems. An AI-based system can perform routine work tasks automatically, it can help humans in their tasks or it may even function independently. A car is a good example of the advances in assisting systems, which may result in an autonomous vehicle functioning without human guidance. New cars already have driver-assist systems, ranging from lane assists to automatic emergency braking. Autonomous vehicles are already being tested in road traffic in many countries.
The car example can be transferred to many other areas of life. An AI system can provide humans with assistance in a broad range of different situations and in some situations it may replace humans. The desirability of full automation depends on a number of practical issues, such as the risks involved in the task. At the same time, it is also a philosophical issue. We must have an open discussion on technological advances. Do we want a situation where AI only serves as an assistant to humans? In what tasks should AI systems act independently?
Even if Google DeepMind’s Alpha Go is able to beat humans at Go, the system does not have the experience that it is playing a game in the way that a human does.
– John Shawe-Taylor
In the future, we will see AI systems that are much more aware of their environment and are able to adapt to change. This will inevitably lead to systems that are much more like humans and are able to adjust to changes around them.
User-friendliness, ease of use and ease of understanding are crucial factors in AI solutions. Users expect more from a ‘smart’ technology than from a traditional information technology. This means that it is also expected to behave in a reasonable manner. If the interaction between a system and humans is deemed as problematic, the system will not be taken into use.
There has been a great deal of research on the interaction between humans and computers in Finland and such research has also been carried out as part of AI research. Therefore, we will benefit if more emphasis is placed on interactive skills in the future.
Artificial intelligence application in the industrial and service sectors
AI technologies are most commonly used in sectors that have access to large and useful data resources. In addition to consumer business and large digital companies, the health and wellbeing sector has also been mentioned as an obvious user of artificial intelligence. There has been particularly rapid progress in the field of medical imaging in which doctors can diagnose findings with the help of AI methods. In the future, AI methods will also be used in the predicting of patients’ condition, estimating of drug response or in the planning of clinical trials.
Robots and intelligent aids will assume an increasingly important role in care tasks where human presence and empathy are not essential. Lifting patients and assisting their mobility are two examples of this.
The introduction of AI and robotics in healthcare also depends on sectoral regulation and the attitudes among professionals, doctors and nurses towards the new technology. Equipment manufacturers and service providers in the sector must ensure that the new technology will earn the trust of both professionals and patients.
Partial AI-based automation of healthcare processes frees the healthcare professionals from office routines to patient and care work. However, healthcare is not the only sector where routine processes can be automated using AI-based solutions.
Manufacturing and processing industries will also introduce AI technologies. Processing industries are understandably cautious with regard to the new technology because introducing new technologies requires detailed plans and the changes are not made overnight. Companies have optimised their production for decades, the lifecycle for investments in manufacturing plants may be several decades and industrial operations are subject to strict safety and environmental regulations.
Small subcontracting companies in manufacturing industries have a specific problem that AI technologies may help to solve. When a company’s employees are ageing and retiring, tacit knowledge is disappearing. The sector does not attract young people. In fact, the disappearance of small companies at the base of the supply chain may threaten the existence of entire industrial sectors in Finland and in many other countries, such as Japan. Application of AI technologies may help to solve the problem in three ways:
- automation will reduce the need for workforce;
- AI may help in the collection of tacit knowledge; and
- a new technology makes the sector more interesting for young people.
AI technologies are also changing the logistics and transport sector as well as agriculture where sensors, drones and artificial intelligence are providing opportunities for precision farming.
It is also believed that AI will substantially change education, training and teaching. As in healthcare, the pace of change will greatly depend on regulation and stakeholders’ attitudes. In creative fields, AI technologies supplement and support existing processes but they are also opening up new opportunities.
Impacts of artificial intelligence on business operations
The speed of introducing AI varies between sectors. There are open source tools available that companies can use for testing the potential of AI in general and machine learning methods in particular. However, there is a long way from tests to actual use. The introduction of AI requires substantial inputs into the development of technological capabilities, into the interaction between users and the technology as well as into the creation of business models. An AI solution can only be made operational if it is compatible with existing processes and operating practices as well as the business model of the organisation.
If 100 companies are trying it and it might be that only 10 of them have a practical business model at this point in time.
– Volker Tresp
In many cases, enhancing operational efficiency is the first stage in the application of AI technologies. This may mean more efficient use of workforce in the care of the aged, reducing the proportion of unsellable products in retail sales or higher fuel efficiency in maritime transport. Substantial savings or productivity improvements can be achieved. These improvements are company-internal in nature.
Industries that already struggle can’t really be saved by sprinkling some AI on top of it, instead we need to go where the new opportunities are.
– Christian Guttmann
In the experts’ view, AI technologies provide a basis for disruptive changes that have a greater impact than improvements in operational efficiency. AI technologies can help to create entirely new business models that do not only have an impact on individual organisations but may also have a wider effect on business ecosystems. For this reason, it is important to stimulate AI-based business operations, for example by expanding cooperation between different sectors and companies of different sizes. The cooperation networks may be on a regional basis but they should also have an international dimension. At the same time, the cooperation networks also help to create attractive environments for talent. For example, the ability to choose any of the jobs offered by a large number of companies is an attractive prospect for foreign AI experts. Experts moving between companies transfer knowledge through ‘¨cross-pollination’ and at the same time, they also help to build the success of all companies in the operating environment.
Even though large global companies dominating the consumer business are already making extensive use of AI, the technology is only now gaining a foothold in business to business operations (B2B). In these sectors, there is often less training material for using AI available and for this reason, they need new AI systems requiring less training material. This can open up significant opportunities for Finland because we have particularly strong research and technology expertise in this area and strong traditions in B2B operations.
Artificial intelligence as a factor renewing society and democracy
At the moment, the focus in the AI debate in Finland and elsewhere in Europe is on ethical issues: protection of privacy, accountability for the errors made by AI systems and the traceability and transparency of algorithm-based decision-making. The members of the expert panel consider these issues important but difficult and add that they can only be solved through international cooperation.
Does the focusing on ethical AI issues slow down progress in Europe, while at the same time China is going full speed ahead with the development and introduction of AI solutions? According to the expert panel, we must find a balance between the protection of privacy and the other rights of the citizens on the one hand, and the benefits generated by artificial intelligence on the other.
Europe should find a balance between privacy and using AI technology.
– Gesche Joost
Should the focus be on the more practical issue of trust? Do we trust the organisations and people producing the AI-based services? Do we trust the technology solutions offered to us?
In practice, deep neural networks are black boxes. This has been criticised and it has been demanded that the algorithms should be transparent. But is this really needed and is it even possible? We can compare this with the functioning of the 4G mobile phone network. Do we actually understand how the calls are transmitted between continents? Even if we did not understand it, it is not necessarily a problem because we trust the data communications system and the parties operating it. Here too, it is a question of trust rather than transparency.
We should be talking about trusting AI just as much as ethics in AI, it is a much productive approach.
– Christian Guttmann
Extensive introduction of AI technologies has an impact on employment, income and (in the longer term) the tax base. It is believed that in the next few years, the automation enabled by AI will mainly have an impact on jobs containing mainly routine tasks. Experts are of the view that countries like Finland that have a high overall level of education will not significantly suffer from the unemployment caused by the introduction of AI solutions in the coming years. It is also essential to support the creation of new jobs rather than protect the disappearing ones.
In the long term (in 20 to 50 years), AI may reach the performance level of humans or even exceed our capabilities in most tasks. As a result, extensive labour participation is no longer required to supply products and services. If paid work as a source of livelihood will become less common, the tax base will erode. This means that we need new means of livelihood and new sources of tax revenue.
AI could facilitate civilized discussion about societal and political issues, it could be the agora of modern time.
– Michele Sebag
There has been a great deal of debate about the role of AI technologies in societal influencing and decision-making. Even though external influencing of voters has given negative publicity to AI and analytics, especially in connection with the 2016 US presidential elections, AI technologies can also be used in a positive manner.
Artificial intelligence can be used to collect citizens’ opinions, identify hostile influencing attempts in social media and develop defences against hybrid threats. Robots and intelligent devices can extend the independent living of elderly people.
Drones can serve remote regions and AI-based ‘trainers’ can advise employees how to work more effectively in teams or they may even help to find solutions to problems in family lives.
China is emerging as an AI power – what will Finland and the EU do?
China is investing heavily in AI technology and its aim is to be the world’s number one in this sector by the year 2030. US companies are still leaders in the development of AI technology and AI-based business but in relative terms, China is investing more in these fields. Access to huge amounts of data gives China a competitive edge. This is a major advantage as long as AI technologies based on a large supply of teaching material dominate AI development. In a centrally governed country, even radical AI-based solutions can be implemented without regulatory or civil rights considerations.
China is investing hugely, but it is also wooing talent to come to China. We do not fully appreciate that.
– Gesche Joost
Even though the matter has attracted little attention in Finland and elsewhere in Europe, there is a fierce competition going on between companies and academic institutions in China, the USA and the rest of the world for top AI talent. If we fail to react to the situation early enough, Finland and the rest of Europe will soon start suffering from brain drain.
Finland should focus on its strengths so that it can retain and attract top talent. Highly motivated research groups focusing on emerging sectors, such as unsupervised learning, a vibrant startup field and close cooperation between research institutions and companies are Finland’s strengths. The Finnish Center for Artificial Intelligence (FCAI) is playing a major role in the boosting of these strengths. Finland also provides families with a safe and clean environment and companies with a predictable legislative and taxation system. These strengths should be marketed in Finland and the rest of the world.
The European Union and its Member States have initiated measures in AI research and development and Finland has been one of the countries driving these efforts. The EU Member States should combine forces to harmonise AI research, development and regulation. By doing that, they can provide companies with a competitive operating environment.
Recommendations of the international expert panel for Finland
- Retain talent
Competition for skilled experts is fierce. Sufficient action has not been taken in Europe and Finland to deal with this issue. Finland needs to not only retain current talent but also to attract more highly skilled artificial intelligence specialists to Finland. This also applies to future talent, in other words students.
- Invest in B2B activities
Artificial intelligence is only making its way to the B2B markets. Finland is well equipped to succeed in these markets, provided it invests in its strengths, which include AI technologies for industrial solutions such as unsupervised learning and hybrid approaches, and a strong focus on industrial B2B business. This development stage provides a good fit with Finland’s industry base and offers significant potential for the Finnish society and economy.
- Respect the principles of democracy and freedom
An approach must be identified in Finland and in Europe that respects the principles of Western democracy and freedom while at the same time permitting businesses, consumers and public services to benefit from AI- based technologies. Solutions based on artificial intelligence should be seen as a way of reinventing society and increasing citizens’ participation in decision-making and democratic processes.
- Artificial intelligence is more than a technology
It is important to bear in mind that technology only provides tools for implementing new business models and better public services. Before any solutions can be implemented, user acceptance is required.
- Embrace the free mobility of data
How useful artificial intelligence is depends largely on the availability of data. Bringing in meaningful data from multiple sources will improve the results dramatically. Therefore, we should break down silos within and between businesses and public services whenever possible and permit the free mobility of data.
CASE K-GROUP: ARTIFICIAL INTELLIGENCE KNOWS BE T TER THAN YOU WHAT YOU WOULD LIKE TO EAT
For many years, the customer loyalty schemes of grocery stores have known everything about their customers. Based on the purchasing data, it is easy to make conclusions about the customers’ buying behaviour. We also know that people usually buy more or less the same things every time when they go to a grocery store.
A few years ago, the K-Group decided to use this data as a basis for a much more intelligent system that would offer recipe recommendations and, in this way, make people’s lives easier. The system was introduced in 2017. In its simplest form, the system works so that when you login and enter the word ‘milk’ into the search field, your favourite milk (based on your purchasing history) automatically appears on the screen.
However, the system was soon made smarter and the search function was expanded to cover recipe recommendations. Using the search results as a basis, the system concludes what recipes are used by people with different food purchasing habits and automatically recommends these to new users.
From the start of this year, the system was made even smarter and it now also covers shopping lists. In practice, when you login, you can now get the shopping list for the following week by pressing a single key. The artificial intelligence application determines the content of the shopping basket on your behalf and it bases its choices on your previous shopping lists. In other words, it does not only recommend products that you have already bought but also concludes what you might want to eat the next week. As you make more purchases and shopping lists, the system will learn more about your habits and will also recommend new products.
In the end, the artificial intelligence application may become a better expert in your preferences and you no longer need to spend time on thinking what to buy next.
The K-Group has collected data on its customers’ food purchases for many years, which means that when recommending purchases and recipes, the artificial intelligence application is on a fairly solid basis. The data is available in large amounts and it is of high quality.
At the same time, however, it has been more difficult to teach artificial intelligence to differentiate between personal preferences and seasonal variations. In other words, even if you loved Christmas pastries, you probably do want to get a recommendation to buy them in June. Accessibility and quality of the data as well as the adequacy of the data mass computing capacity have been the key challenges and only in the past few years has progress been achieved in these areas so that relevant recommendations can be given.
The next aim is to make the recipe and shopping list recommendations into a workable package that would revolutionise grocery shopping. You would only need to press a few keys to select and purchase the items. In this way:
The system asks you whether you would like to make a casserole and vegetable pasta next week. You answer is yes, after which the system suggests products for your shopping list. You accept the products after which the system asks whether you want to buy the products. You approve the list and pay the products, which will be delivered to your home address a short time later.
This may already become a reality in one year’s time.
The recipe recommendations of the K-Group are an example of the development of digital services. This is a factor strongly impacting our daily life: we spend a great deal of time on grocery shopping and on thinking what we would like to eat. This means that artificial intelligence may revolutionise our daily lives by minimising the time spent on that activity and by learning to determine on our behalf what we have eaten and what we should eat next.
CASE FIVA: MAKING ROUTINE WORK EASIER WITH ARTIFICIAL INTELLIGENCE
In spring 2018, the Financial Supervisory Authority (Fiva) launched a pilot project, in which the aim was to substantially reduce the amount of manual and routine work.
After analysing the potential of robotics and artificial intelligence, Fiva implemented the first pilot project in investment service notifications. The task in which it was decided to test the robot requires a great deal of effort and time when done by humans and there is also substantial potential for human error. In this task, information is entered into a database, which means that theoretically, a robot could be programmed to perform the same work. The information concerns the investment services offered by European companies and the authorities use notifications to exchange information.
The robot processing the investment service notifications was also taught to process emails. The robot reads the email attachments and based on their contents, it can create new companies in the supervision register or update the change information contained in the notifications. It also enters dozens of different types of licence and service information into the register. As part of the process, the robot sends emails to other units in the organisation requesting registration information and enters this information into the register.
Encouraged by the excellent results of the first pilot, Fiva decided to carry out a second pilot in fund notification, which is slightly different from the first project. Fund notification documents involve substantially more variation than the documents
in the first pilot, which means that the task could not have been managed by a robot alone. For this reason, an AI application that is able to process free-form text using statistical context recognition was developed to assist the robot. A simple AI algorithm is able to make a distinction between these different types of document.
The artificial intelligence application is able to process documents in Finnish, Swedish and English. In addition to classifying documents, it is also able to pick specific information irrespective of where it is located in the document. The AI-assisted robot for fund notifications was introduced in February 2019.
The results from the use of the robot and the AI-assisted robot have been overwhelmingly positive. The fact that European supervisors do not fully observe the jointly agreed notification procedures has been the main source of additional work in the pilot projects. It is clear that the software robots achieve optimum results when all data is of uniform type, free of errors and seamless.
As the robots used in the pilot projects have been taught to imitate the work of humans, it has been asked many times during the project which systems the robot should be able to access and what IDs and access rights it should be provided. Policy decisions on these issues have been reached during the project. For example, it has been agreed that users are able to check the notifications processed by the robot and that the robot sends all emails with its own name.
As a result, the work of the experts at Fiva has become more meaningful and most importantly, the robot and the AI have reduced work backlog. The robots also perform tasks in which humans make substantially more errors, which means that the number of errors has also decreased. Fiva will decide on the future action after the recently started second pilot has been completed.
The Financial Supervisory Authority is an example of a government agency that uses artificial intelligence. The Financial Supervisory Authority is a financial and insurance supervisory agency, which does not always have adequate resources for its tasks. Robotics and artificial intelligence have eased the situation and their use has also led to a sharp fall in the number of human errors. The Financial Supervisory Authority supervises the operations of banks, insurance and pension insurance companies, other actors in the insurance sector, investment firms, fund management companies and the Helsinki Stock Exchange. The supervised entities provide 95% of the funding for the operations and the remaining five per cent comes from the Bank of Finland.