European AI Alliance

Archived

ARCHIVED PLATFORM

This website is archived since mid-May 2021 and it is not possible to login anymore.

But don't worry there is a new Futurium platform and all active communities have been migrated.

Forum

HLEG - Input Request: World Class Research

Fredrik HEINTZ

9 November 2018 - updated 2 years ago
16 comments
Total votes: 6

Sami Haddadin and myself are co-rapporteurs in the AI HLEG for the group working on « World Class Resarch ». To improve our section we would like to have your input on a number of questions. Feel free to reply only to some of them. Short messages and longer thesis will be appreciated equally.

Specific examples of the benefits of AI to the citizens where research is needed
Specific examples of where the lack of data has been an issue for doing research
Examples of successful industry-academia-government research collaboration models
Specific and concrete research topics (short, medium and long term)

Thanks!

Fredrik

Comments

Marco Bertani-Økland,

15/02/2019 22:05

1./4. Explanations in machine learning are good, but they usually are static and not actionable. The following paper introduces actionable recourse, where you are given actionable input variables that let you know what you need to do to change the decision for instance of a loan approval system.
https://arxiv.org/abs/1809.06514

This is for linear systems, and a natural generalization is to extend this to monotonic models or even non linear models. Look at the examples of the article to fully understand the importance of this issue.

At the same time, we need more research in how can one use explanations to enhance human performance in decision making. How to best integrate algorithms and humans. This allows for faster case management in systems that are critical to society and where full automation is not desirable.

Another area of research is fairness in machine learning. The industry needs best practices in how to map your moral principles of equality to concrete algorithmic implementations. We know there are many definitions of fairness and many are not even compatible with each other. How do we choose from these?

2. In fairness research, there are few datasets to explore and do research, and we know that these fairness metrics are highly context dependant. Ergo we need more varied datasets that can help us understand what is going on here.

3. A good example of good collaboration between industry, research organizations and government, is the ACM FAT* conference, where people from these areas come together and talk about how to build systems that don't discriminate.
https://fatconference.org/2019/

Fredrik HEINTZ,

17/02/2019 12:23

Thank you Marco,

I agree that explainability should help people understand what they can do in order to influence the outcome. A special thanks for the refernce!

Regards,

Fredrik

Vladan DEVEDZIC,

15/02/2019 21:23

Hi Fredrik, All,

Here are some of my thoughts with regard to the questions you put up:

1. Specific examples of the benefits of AI to the citizens where
research is needed

In order for AI to be fully trustworthy, it must be explainable.
The ML hype today is fine, but ML usually lacks automated explanation.
It can leave end users puzzled.

2. Specific examples of where the lack of data has been an issue for
doing research

All areas where regulations such as GDPR restrict the usage of data. Take, for instance, pre-school education (a very sensitive area).

4. Specific and concrete research topics (short, medium and long term)

XAI, XAI and XAI :)

Regards,

Vladan

Norbert JASTROCH,

14/02/2019 21:22

Dear Fredrik,

my suggestion is research into

(1) the management of complex systems and

(2) the interrelation between knowledge and innovation.

Regards, Norbert

Kristof Kloeckner,

17/02/2019 17:11

I agree that these two topics have high relevance for many domains, for instance for IT Services or the Internet of Things. At IBM Global Technology Services, we have applied AI Technologies to the service lifecycle, with good results but some unique challenges.

(1) One of the big challenges e.g. of incident management in complex IT systems is separating signal from noise, determining what are 'true, actionable' events and determining the best remediation. Data is generally very noisy, and a balance needs to be struck between outright automation and supporting subject matter experts to resolve the most complex problems. All approaches require a taxonomy of the problem space, and this is directly related to (2).

(2) Managing the lifecycle of organizational knowledge is a key challenge for all knowledge-based organizations, in our case the community of IT experts. This has both technological implications (like the role of ontologies, standardizing ontologies and federating them across domains), as well as social ones (expert engagement, social curation of knowledge). Acceptance of AI generated advice is directly related to transparency or 'explainability'.

We have published our insights in a short monograph ("Transforming the IT Services Lifecycle with AI Technologies")

https://www.springer.com/de/book/9783319940472

If used in the right way, AI can have a huge impact on the two areas you point out, but much more research is required. We are still at the beginning.

Francesco Saver...,

14/02/2019 19:20

Dear Fredrik,

Your questions are very interesting and valuable,

I can suggest two interesting research domains, question 4:

Misinformation, using IA to tackle FakeNews and misinformation, as you know this is an hot topic in EC at the moment; I think this is also a new interesting topic at scientific level, including synergies in many different domains: IA, but also data science, complex system physics, social network and social media studies
IA and 5G, this is a valuable research domains, with deep business impact, in two directions how use IA to support network configuration in 5G; but also how 5G can support IA (with a strong usage of edge computing to solve computational problems at local level)

Naturally, both aspects need more investigation and, even, more space to be explained better; but for the moment these are my 2cents

Ciao Francesco

Shalom BROYER,

03/12/2018 18:03

Hi,

We are trying to predict some labels in the area of palntation (Avocado, Mango). My answers will refer to this subject.

1. Specific examples of the benefits of AI to the citizens where research is needed

If we could provide future information on outputs (size, quality) for fruit growers it would help them better plan the distribution to the market and negotiate on prices.

2. Specific examples of where the lack of data has been an issue for doing research

We experience a problem when we want to ןמאקערשאק information from different sources, such as soil type, weather, irrigation , size of the plot and yield for 20 years.

3. Examples of successful industry-academia-government research collaboration models

None

4.Specific and concrete research topics (short, medium and long term)

We are trying to predict some labels in the area of palntation (Avocado, Mango). My answers will refer to this subject.

Pınar Barlas,

15/11/2018 17:43

1) A benefit would be to have automated decision-making that is completely objective, devoid of all social biases we humans have. (It's often assumed this is already the case - but unfortunately human biases seep in through everything from choice of training data to reinforcement to judgement of outputs.)

4) Further research around bias in algorithmic systems is necessary. What kinds of bias can appear in which systems? How do we detect it? How do we minimize or eliminate it?

My current research looks at bias in computer vision systems. This brings me to:

2) Training data and a full list of possible outputs are often kept secret from researchers and the public. That aside, even things like whether the computer vision service you hire is updated continuously, updated once every X months, or not updated at all are not disclosed. It's understandable that these types of information are proprietary and kept secret from competitors, but there must be a mechanism in place so that these systems (which are the basis of many security and other high-importance applications) do not go unchecked.

Fredrik HEINTZ,

02/12/2018 17:28

Thank you Pinar. Detecting and mitigating bias in data is an important research topic. Whether truly bias-free data/systems are possible, I'm not sure about. In any case, removing unwanted and unintended bias is crucial. There are also interesting questions related to the normative aspects of di-bias data. Should data represent the true world or the world we want? In the later case, who should decided which world we want? What if different people/groups want different worlds?

The questions of accountability, traceability and transparence are important and are covered at great length in our work (but not explicitly in the part on World Class Research that I am working on).

Andres Abad Rod...,

13/11/2018 11:52

Dear Fredik,

there is an ITU and WHO effort in AI4Health where some interesting examples are discussed. There were some interesting examples discussed: image recognition for snakes bitting, etc... I can put you in contact with people there.

There were many people attending from organizations (ITU; WHO and other UN agencies), universities (Switzerland, Croatia, Germany, etc) and different private companies.

Fredrik HEINTZ,

02/12/2018 17:16

Thank you Andres, I agree, health is really important and to coordinate with existing initiatives is important to get the most out of the work.

Kai Salmela,

09/11/2018 16:31

Hello , Mr Heintz

To the first topic, i'd like to offer linquistic research, especially for non indo-european languages - i.e. Fenno-Ugrian language group. It has great problems to be translated into a form where machines could understand it. Algorithm seem to be still too complicated and real translation or speech recognition just does'nt exist yet, just some tryings where system recognise word every now and then but cannot make comprehensive statements.

This problem has resulted that all Finns have to learn English in order to be digitally sawwy. Older people and those who cannot english are left out of the digital society because of the language barrier.

I've been told that if the proper algorithm is invented for Fenno-Ugrian languages, it would also step up speech recognition in all languages since after that machines could understand meanings behind the words.

I hope this topic offer is what you sought for ?

wbr

Kai Salmela , AI Specialist in Robocoast R&D ( DIH )

Fredrik HEINTZ,

02/12/2018 17:13

Thank you Kai. I completely agree with you. Smaller languages, including my native Swedish, are important to European research. We have included the following opportunity "The multi-cultural, multi-language reality of Europe provides an ideal opportunity to explore alternative methods and theories for learning and perception that focus on small data and on the semantics of data."

Best regards,
Fredrik

Laurentiu VASILIU,

09/11/2018 16:23

1. The combination between Robotics and AI: *real* AI driven mobile robots/drones to assist with care for the elders, public security monitoring and enforcement, children supervision and education/teaching. Mobile robots these days are just very expensive showcase toys, and not because of the hardware but because of the primitive software that drives them as well as the lack of *real* AI (and not 'marketing AI' of partially hard-coded behaviours and niche applications that we see everywhere now)

2. AI usage in the health area: data access in the medical field is highly scattered, broken, missing or incomplete and highly bureaucrating to access. To test an AI application on medical data is extremly difficult/close to impossible compared to say testing AI on financial data.

3. -

4. Research topics: Deep learning, reasoning, complex decision making, optimisation - these are all long term ones where the research fronteers need to be pushed.

Fredrik HEINTZ,

02/12/2018 17:20

Thank you Laurentiu, yes, AI-robotics is a fruitful and important research area. In the current version of our document, this is lifted as one important and strong area for Europe.

Bjoern Juretzki,

06/12/2018 09:20

Shouldn't we look one level deeper and consider specific instantiations of robotics? Europe is not equally strong in all areas and application domains of robotics, but has strengths, for example, in industrial robotics, agriucultural robotics, healthcare and maintenance and inspection. The more concrete we are, the easier it is as well to demonstrate the value to citizens (agricultural robotics: reduction of pesticides, healthcare: more time for patients, etc.).

ARCHIVED PLATFORM

This website is archived since mid-May 2021 and it is not possible to login anymore.

But don't worry there is a new Futurium platform and all active communities have been migrated.

Share this:

Comments