Part IV: Application domains and new paradigms

In this part, we address the following research topics:

— Presentation of several application contexts where there is an interaction between human and machine intelligence and new future paradigms in computation.

— Presentation of research areas that will have a future impact on how we understand machine intelligence.

16 Machine learning in healthcare and computer-assisted treatment

Miguel-Ángel González-Ballester

ICREA and Universitat Pompeu Fabra

16.1 Statement

Machine learning is in a phase of renaissance that is transforming practices in multiple fields. Beyond the application of previously existing techniques, novel developments in big data and deep learning have transformed the landscape of available methodologies. Furthermore, many of these developments are spearheaded by industry, which is leading the application of machine learning to everyday products and services, not least in healthcare.

Image processing, in particular, has suffered a revolution in the last few years. Current

methods based on deep learning clearly outperform the previous state of the art, and their extrapolation to medical image analysis has shown very promising results in i.e. lung cancer detection.

Computer aided diagnosis is also benefitting from recent developments in deep learning and machine intelligence, particularly in enabling the analysis of large, heterogeneous sources of patient data, such as genetic tests, blood and cell samples, imaging explorations and unstructured information from the clinical history of the patient. Furthermore, these tools are also being applied to study the aetiology of complex diseases, by finding patterns in large patient databases.

Despite these impressive initial success stories in medical image processing and computer-aided diagnosis, strong limitations have become apparent. Modern machine learning is predominantly based on “black box” approaches, failing to provide reasoned interpretations of the diagnoses they provide. Doctors (and we are far from replacing them) cannot afford to incorporate tools that provide a diagnosis with no explanation about the reasons behind this diagnosis. This poses a number of challenges for the widespread use of machine intelligence in healthcare.

16.2 Future challenges

• Interpretability is key to the future success of machine learning and artificial intelligence in healthcare. The combination of data-driven (empiricist) and model-based (Platonic) approaches might be key to this end.
• The availability of medical data is often limited by ethical and regulatory issues. Current trends in data augmentation and generative networks partly help in increasing the numbers of available data, but they risk biasing databases with unrealistic (non-disease related) information.
• Embodiment of artificial intelligence for patient care, e.g. through surgical robots, robot companions for the aged, or pervasive access and monitoring of health information, is an emerging discipline that may revolutionise healthcare.

17 The influence of “intelligent” technologies on the way we discover and experience music

Fabien Gouyon

Pandora

17.1 Statement

There is an influential loop between technological innovation, the development of business models governing the music industry, and the way we discover and experience music.

Internet and internet music streaming transformed the music industry. Streaming is now the primary way we listen to music, and this transformation is only at its beginning.

Until recently, the way we listened to recorded music was tightly linked to a relatively clear business model. Namely, music discovery would be driven by diverse media (e.g. terrestrial radio), targeting a subsequent purchase and ownership of a physical —or digital— artifact of what was discovered, consumption being done via another media (e.g. personal CD-player, iPod, etc.). A whole industry (a multi-billion dollar industry) was based on that model, which is now put under pressure.

Under this ownership model, once the discovery phase happens and an item is purchased, the job of content creators, producers and distributors is basically done. It is the listener who decides how and when to enjoy their music, with little influence from who produced or distributed it.

Now, with the advent of streaming, we are witnessing a shift from ownership to access. And with the access model, the line between discovery and consumption is now blurred, as the same media now serves both. There is an opportunity for content producers and distributors to guide listeners in their consumption. This opens the way to a much more holistic experience. And in return, the listener now requires to be assisted in all aspects from search, discovery, browsing, sorting through enormous collections of tracks, consumption, sharing, etc. That leaves room for many different novel models of listening experiences.

This is precisely where a crucial part of the music streaming industrial competition is currently happening. Diverse companies are developing at great speed new products, such as personalized playlists, radio-like lean-back propositions, etc., aiming at defining new formats of music listening. This calls for developing new technologies for recommending the most relevant content, as well as the most relevant vehicle for content discovery and consumption.

Such technologies must have the potential to be personalized for all listeners and reactive in real-time, and they must balance many factors such as e.g. content repetition, interactivity, or user intent.

In other words, the current developments in the music industry that are primarily driven by technological innovation are shaping the way hundredth of millions will access music, experience it and socialize around it. It is therefore fair to say that researchers and technological companies alike should acknowledge their strong cultural and societal responsibilities, and even further, embrace them.

Let’s consider a few examples. For instance our responsibilities with the listener: The ubiquitous availability of (almost) any artefact of the world's music repertoire only a few clicks away imply overwhelming choices to music lovers. They need assistance, and we have a responsibility in helping them navigating through, and filtering this flood of content. Recommendation and personalization technologies can help in this endeavour. But they also are prone to potential algorithmic biases, and can result in a progressive isolation of users in their own musical bubbles, hence limiting and ultimately hurting their experience.

Let’s now consider responsibilities with the music ecosystem, which we are part of: The fact is, music distribution is extremely unbalanced: a very small proportion of artists (the “head” of the distribution) account for most of what’s listened to, while the large majority of artists (the “tail”) remain listened by few. There are hits, and there are niches. We should carefully consider the effect technology can have on this distribution. It is (relatively) easy to develop technology that could —directly or indirectly— have a favourable distribution impact on either the “tail” or the “head”. These would likely result in different economic returns for the industry and for the artists, on the short-term and on the long-term.

Finally, let’s also consider that technological innovation is not only influencing the music listener experience as exemplified above, but it is in fact currently revolutionizing most aspects of the music industry, from creation, to rights monitoring, marketing, monetization, etc. This could spark similar reflexions on our cultural and societal responsibilities as developers of these technologies.

17.2 Challenges

• How, in the development of innovative technologies, can we exhaustively identify, understand and deal with algorithmic biases?
• How can we devise metrics that would approximate long-term user satisfaction?
• How can we make algorithms (e.g. recommendation algorithms) more transparent to their users? And what degree of transparency is actually desired/required?
• How can we balance ever more adaptive, contextual user experiences and respect for privacy? How can we provide users with more control on the data they provide us?
• What level of trust between a user and a technology/service is desirable to achieve? How far can the interaction go, and are there limits to be fixed?
• In the current context of extremely fast-paced scientific and technological developments, a very competitive and dynamic music streaming industry, and the "all-you-can-consume” nature of modern media, how can we help users cope with the overwhelming flood of content and products, and help them engage more deeply with content?
• What should be the basic methodological steps to follow so that the novel technology we develop not only follows to the latest technological trend, and responds to business metrics, but also empowers its users in its very evolution?

18 HumanAI

Blagoj Delipetrev

Digital Economy Unit, Joint Research Center, European Commission

18.1 Statement

Human and Machine intelligence comparison

There have been millions of year of biological evolution. Almost three billions years was needed for evolution to create the current human being. Evolution has increased human brain size exponentially over the last 8 million years from below 250 cc to 1500 cc. The brain is the source of our intelligence and separates us from all other species.

The industrial revolution, 2-3 centuries ago, produced machines that replaced physical human labour. The computational machines started 70 years ago with the semiconductors and their exponential grow which leaded to the creation of artificial intelligence (AI) which replaced humans in cognitive tasks.

The most important point is the time scale. Both human and machine evolutions are exponential and while human evolution is in millions of years, the machine evolution is in decades. There is a distinctive difference between human intelligence and the current AI, but this may not be the case in the future. Moore law is not dead, and the machine rise continues.

Deep Learning

Deep learning (DL) is the flagship of the AI research and achievements in the last decade. Increased computational power and vast amounts of digital data are foundation for DL, which is in essence a multiple layer neural network. DL achieved many breakthroughs, starting from vastly improving image recognition, NLP or translation. The DL combined with Reinforcement learning (RL) won the game of GO, Atari, Poker and lastly Dota. DL and RL are rapidly expanding.

Algorithmic Impact Assessment

The AI fast pace produced many systems that are used in everyday lives, government and public offices. There is a need of validation of these AI systems. Worldwide various initiatives for algorithm impact assessment will evaluate and analyse AI systems and their decisions and make them more transparent, understandable and explainable.

Predictions

My predictions are that “everything electrified will be cognified”, “AI is the new electricity” making devices more intelligent and autonomous. Tasks described in productivity and efficiency will be performed by robots and bots. Now and in near future AI is going to complement us in all our daily tasks, as they do already, with our smartphone, computers, etc. In the middle term, the more advanced and intelligent machines will completely replace humans in tasks like transportation, medical image recognition, language translation, etc. In the long term is possible to have AGI or something close to it that will be capable of doing multiple cognitive tasks better than human does. This does not mean that humans will be obsolete.

In the meantime, we need to address many current problems and possible AI dangers. One of the most vivid dangers is autonomous weapons. Other highly important topics are the rising inequality, unemployment, fairness, inclusion, social justice, etc.

18.2 Challenges

The world is on the verge of one of its most valued discoveries AI. There are huge benefits in rising productivity, efficiency, improved standard, longer lifespan, better healthcare, etc. AI will automate most of the tasks, leaving more time for humans to

enjoy lives and be creative. AI can bring more happiness and prosperity but also dead and destruction. Therefore, there is a need for AI regulation for the benefit of all humanity.

References

— “Life 3.0: Being Human in the Age of Artificial Intelligence” by Max Tegmark

— “Sapiens: A Brief History of Humankind” by Yuval Noah Harari

— “How to Create a Mind: The Secret of Human Thought Revealed” by Ray Kurzweil

— "Deep learning." nature 521.7553 (2015): 436. LeCun, Yann, Yoshua Bengio, and Geoffrey Hinton.

— “How AI can bring on a second industrial revolution” by Kevin Kelly https://www.ted.com/talks/kevin_kelly_how_ai_can_bring_on_a_second_industrial_r evolution

19 Do humans know which AI applications they do need?

Perfecto Herrera

Universitat Pompeu Fabra and Escola Superior de Música de Catalunya

In the last century, humans have developed powerful technologies that, for the first time in history have the potential to quickly and irreversibly change the world as it was previously known. Nuclear power and genetic engineering are application areas derived from useful essential knowledge (which cannot be questioned or censored) that had to be subject to ethical and legal regulation, even by means of international agreements. The current state of our knowledge on AI makes some of their applications to be about to cross (or maybe already crossing) red lines too and there have been attempts to reach a consensus, at least in the scientific community (see for example: https://futureoflife.org/ai-principles/ or http://www.iiia.csic.es/barcelonadeclaration/).

In the panel on the application domains of AI that closed our kickoff-meeting we witnessed some of such concerns, but also other ones that have to do with our concepts of humanity or creativity.

Sergi Jordà, in “Enhancing or mimicking human (musical) creativity? The bright and dark sides of the Moon” debated on attempts to develop “creative machines” and how many of them cannot shed enough light on human creative or other cognitive processes (one of the goals or justifications of some AI practitioners). In addition, creative systems are usually deprecated by their potential users (flesh-endowed music creators), not to mention the shallowness or uninterestingness of their outputs (though some outstanding exceptions could be considered). This rejection of creative systems could be due to a narrow developing perspective that does not consider the human to be inside a loop with the AI system. The idea of computers as assistants, becoming extensions of their users and doing the “dirty” or the short-time unfeasible work, should be promoted and researched (instead of leaving them the option to make the serious decisions). This is something that Luc Steels, another of the panel participants, commented during the panel dialogue (“Intelligence amplification” was the short-name given to that). A final issue with creative systems, but also with other AI devoted to “practical” problem-solving is that of understanding the outputs and the inner workings leading to them, from a human perspective, as some participants also remarked with the special case of game-playing AI systems.

The idea of assisting humans when dealing with creation is also challenging when they are not the creators but they enjoy an artistic creation, as Fabien Gouyon in “The influence of intelligent technologies on the way we discover and experience music” remarked (see Section 17). The almost permanent connection we listeners currently have with music, as a stream passing by or where you live immersed into, calls for ways to improve such listening experiences in “intelligent” ways (recommending truly relevant titles, helping to navigate through options, personalizing musical experiences, connecting with other people or groups, etc.).

Biases already noticed in other types of recommenders, and the potential construction of isolation bubbles on the side of users, should be carefully counter-acted in order to keep the collective and transformative power of music at its best.

A very different perspective and application field was discussed in “Machine learning in healthcare and computer-assisted treatment” by Miguel-Ángel González-Ballester (see Section 16), where advantages and shortcomings of current medical AI systems, especially those devoted to image-based diagnosis, were discussed. Here, again, the requirement that machine decisions can be interpretable under human (professional) criteria was raised. Additionally, the possibility that diagnosis might be done without humans in the loop (or that machines could bias or override the view and expertise brought by them) has also to be considered. An emergent topic, for which we are probably unprepared yet, was AI embodiment (i.e., what happens when humans incorporate, as body parts, AI systems?). Intersecting several of already mentioned hot

issues, Blagoj Delipetrev, in “HumanAI How to assess algorithmic impact?” compared human and machine intelligence remarking what should not be considered as such (brute force approaches) and ways humans and machines could collaborate for a human-favourable scenario (see Section 18). He also discussed applications in satellite image recognition and on the assessment of the impact of AI algorithms. The necessity to deal with natural language and concept generalization before claiming “intelligence” for many AI systems was relevantly remarked there.

Our last talk, “Will AI lead to digital immortality?”, by Luc Steels, speculated on future or “futuristic” applications and issues. Personal assistants are currently being developed under different appearances, and they will probably become autonomous artificial personae that might even impersonate different humans at the same time, not to mention that they might have multiple parallel lives, or that they will become immortal and then can continue with functions attributed to their formerly-assisted human beings. A list of ethical issues open by that perspective was barely touched, although, compared with those that the current intensively-used systems pose, they could be left for some future HUMAINT version 2.

During the open debate other important topics were lightly touched such as the apparent overabundance of AI-based start-up companies without clear business models (which can contribute to the “hype” of the topic), the risk of being monitored in concealed ways (by sound recording devices intended to play with toys or to just receive commands), the direct manipulation of behaviour that recommenders or notification services induce in their users, the blurring between what we thought reality is (or was) and what our senses are processing, the losing of some skills (some of them considered to be inherently human, such as caring for other beings, for example), or the consented (or worryingly ignored) externalization of some of our decision-making processes.

Applications of AI could sometimes be perceived as harmful because of their apparent extraordinary or superhuman “intelligence” but their most worrying aspects should be watched elsewhere: the difficulties to track or explain their decisions, the difficulties to embed AI systems with some moral sense or the subtle or blatant invasion of privacy they can facilitate. Even though there are enough examples of AI systems contributing to a healthier and more pleasurable existence (i.e., diagnosis systems, helpers for autistic or elderly people) and that such systems can help us to cope with data overload and strive in an increasingly complex reality, we should be watchful for some forthcoming large-scale disruption on the way we see ourselves and the surrounding world.

It’s time to decide how our future should look like, instead of leaving it to be devised just by what technology makes possible.