The 27th International Joint Conference on Artificial Intelligence and 23rd European Conference on Artificial Intelligence Stockholm, Sweden. July 16 - 19, 2018.
I had the great pleasure of taking part in Ijcai, Ecai 2018 (The 27th International Joint Conference on Artificial Intelligence and 23rd European Conference on Artificial Intelligence. Stockholm, Sweden. July 16 - 19, 2018).
Below you will find impressions from the conference, and links for further reading.
The Ijcai, Ecai 2018 conference was held at the Stockholmsmässan, in Stockholm, Sweden.
Tried to follow as many talks as possible. But, well, these notes are, of course, in
no way, shape or form complete...
Rather, these notes were written on conference nights, as my way of
keeping track of the events that I attended at the conference. And as a way of storing links and references for future reference.
But enough disclaimers, below, you'll find impressions and links from some of the conference talks and seminars,
including links for further reading.
Great stuff indeed. And much (AI stuff) to look forward to in the coming years!
Ijcai, Ecai 2018.
Stockholmsmässan, Stockholm, Sweden.
16 - 19 July 2018.
Stockholmsmässan, Stockholm, Sweden.
1.1. Page Overview. - Sessions and Keynotes.
Below, in section (2 - 5), you will find impressions and links from sessions, demos and keynotes that I followed Monday, Tuesday, Wednesday & Thursday.
In section (6) you will find a photo montage from this years
World Computer Chess Championship (Icga 2018, part of this years Ijcai conference). Section (7) gives a few impressions from the exhibition hall at Ijcai 2018. Section (8) shows some of the interesting posters, upcoming events etc. I spotted at Ijcai. Section (9) wraps up the week.
Please notice: These notes don't do justice to
the often brilliant presentations that initiated them! So, please read
the original presentations to avoid any distortions ...
2. Impressions from Monday, July 16th.
2.1. ''Victoria'' - track.
2.1.1. Xiaojuan Ma, Hong Kong University of Science and Technology.
Talked about ''Human-Engaged AI''.
And, fascinating, ''what strategies AI systems can employ to engage users in an appropriate manner''?
Sure, a robot might try to get a humans attention through social signals such as gaze, body and head orientation, nodding etc.
But things begin to be ''interesting'' when we go from submisssive robots, to robots who tell
humans that ''I haven't finished yet, turn to me, and follow my instructions ...'''.
Or what about chairs that tell users, when they (humans) don't sit correctly on the chair...?
How should such chairs address their human users?
2.1.2. Jun Zhu, Tsinghua Lab of Brain and Intelligence.
Talked about ''Probabilistic Machine Learning''.
Where noisy or ambiguous data makes it harder for intelligent systems to operate in the real world.
Still, a combination of deep neural networks and probabilistic modeling, Bayesian Deep Learning (BDL),
here in the form of a probabilistic programming library named ZhuSuan, might give a promising way forward.
2.1.3. Christopher Amato, Northeastern University, Boston.
Talked about ''Decision-Making Under Uncertainty in Multi-Agent and Multi-Robot Systems''.
Where things quickly become pretty complex in the real world. I.e. drones might only have one downward facing camera, that allow them
to reason about where they are. Add to that concerns about battery-life, the position of other
drones etc. and it quickly becomes clear that even relatively simple surveillance tasks are actually pretty complex.
Sure, (off-line) deep learning with a lot of examples will solve everything might solve many of these problems, but it still might not be the full answer here.
Especially, as efficient online learning remains a challenge.
2.2. Interactive, Collaborative Robots, by Danica Kragic.
Largely preprogrammed for their tasks, and most commercial robot applications still have a very limited ability to interact and physically engage
In other words, something as simple as cutting bread is still challenging for robots.
And there are still no robots that can wash the dishes.
Something as simple as a good grasp (to quickly take something in your hand and hold it firmly)
is not simple, if you are a robot...
Most robots today, still, only have 2 fingers (grippers) to work with.
Many robots generate motion strategies without any sensory feedback (for simple manipulation tasks),
which obviously only get them so far.
Still, adding visual and haptic feedback will probably allow robots to better understand object shapes, scene properties and in-hand manipulation.
But it is of course still not easy for robots to understand how pushing an object will change the
scene in cluttered environments.
Learning and collaboration might come easy to humans. But robots are obviously not quite there yet,
even though robots that are better in integrating motor and sensory channels might eventually give
us much smarter robots... Still, there is a lot to learn...
But, certainly, a great talk about recent progress.
2.3. AI and Software Engineering.
2.3.1. Jian Li et al. , Chinese University of Hong Kong.
provide a set of helpful services to accelerate software development.
Intelligent code completion is one of the most useful features in IDEs, which suggests next probable code tokens,
such as method calls or object fields, based on existing code in
the context. Traditionally, code completion relies heavily on
compile-time type information to predict next tokens.
But there has also been some success for dynamically-typed languages, where researchers
have treated these languages as natural languages, and trained code completion systems
with the help of large codebases (e.g. GitHub):
In particular, neural language models such as Recurrent Neural Networks
(RNNs) can capture sequential distributions and deep semantics.
Here the authors propose their own system, based on improvements on existing techniques.
A system which turns out to be quite effective.
Talked about ''Positive and Unlabeled Learning for Detecting Software Functional Clones with
Where ''clone detection'' is a vital component in todays software development, in order
to avoid having copies of the same code in several places of ones program.
Detecting pieces of codes with similar functionality (Ususally, created by reused code after copying,
pasting or modification of existing code), but looking (slightly) different is not so easy though.
Here, the authors suggest that clone detection task could be
formalized as a Positive-Unlabeled problem, and indicate that they had some success with it.
2.3.3. Krzysztof Krawiec et al. , Poznan University of Technology, Poland.
Genetic programming is an effective technique for inductive synthesis of programs from tests, i.e.
training examples of desired input-output behavior. (But) Programs synthesized in this way are not
guaranteed to generalize beyond the training set...
The authors then describe an improvement over such a (too) ''simplistic'' Genetic Algorithm approach.
An improvement, that synthesizes correct programs fast, using few examples.
In the authors word:
This work may pave the way for effective hybridization of heuristic search methods like GP with spec-based synthesis.
3.1. The Moral Machine Experiment, by Jean-François Bonnefon.
Jean-François Bonnefon (Toulose School of Economics) talked about The Moral Machine
(a platform for gathering a human perspective on moral decisions made by machine intelligence, such as self-driving cars).
E.g. should an ''ethical'' car sacrifice its own passengers to save (more) pedestrians in case of
People tend to say yes.
But also indicate that they wouldn't like to buy such a car...
Often, such classifiers will try to classify texts into categories like joy,
surprise, disgust, sadness, anger or fear. But it is all a little bit tricky, of course, as a single sentence
can evoke multiple emotions with different intensities. But here a solution, with a convolutional neural network for
text emotion analysis, seem to perform favorably compared to other proposed approaches.
3.2.2. Xin Li et al. , Chinese University of Hong Kong.
Talked about ''Aspect Term Extraction with History Attention and Selective Transformation''.
Where ''Aspect term extraction'' is all about detecting users
opinion and locating opinion indicators.
Classification models based on e.g. SVM's have had some success, but the
authors new framework seems to be able to categorize texts more accurately.
Clearly, a pretty interesting result.
3.3. Model-free, Model-based, and General Intelligence, by Hector Geffner.
Geffner started this super interesting talk by introducing us to ''learners'' and ''solvers'' (and planners, which are a particular
type of solvers), addressing the similarities and differences
between learners and solvers, and the challenge of integrating
''Learners'' (deep learners and deep reinforcement learners) have been part of
most AI success stories in recent years (image understanding, speech recognition, games etc).
Learners require a lot of training (along with an error-function, in the case of supervised learning) in order to become experts within a certain field.
But once they have been trained, the learners are very fast.
Whereas, solvers deal with new problems, and always have to think a little bit (longer) before they are able to
solve a problem.
In both deep learning (DL) and deep reinforcement learning (DRL), training results in a function
f that has a fixed structure (given by a deep neural network).
In both DL and DRL, the most common algorithm for minimizing the error function is stochastic gradient descent where the parameter vector
θ is modified incrementally by taking steps in the direction of the gradient.
Reinforcement learning. An agent takes an action in an environment, which gives a reward that is fed back to the agent.
Solvers take a convenient description of a particular model instance and automatically compute its solution.
For a classical planner, the inputs are classical planning problems and the
output is a plan that solves the problem.
Learners require training, which is often slow, but then are fast; solvers can deal with new problems with no training
but after some deliberation. Solvers are thus general, domain independent as they are called...
... provided a suitable representation of the problems; learners need
experience on related problems.
Both systems have limitations, where:
A key restriction of learners relying on neural networks is that the
size of their inputs x is fixed. This implies that learners cannot emulate solvers even over specific domains.
Still, the speed of the learners is obviously a good thing. Just as
explanation and accountability through models (solvers) is a good thing.
So, we need both.
And, moving forward, we need to better integrate these two systems, within one system:
The solutions to some of the bottlenecks faced by each approach, that often have to do with
representational issues, may lie in the integration that is required to resemble the integration of Systems 1 and 2 in the
An awesome talk!
3.4. Evolution of the Contours of AI.
3.4.1. Wofgang Bibel, Darmstadt University of Technology.
Talked about ''A Scientific Discipline (Once) Named AI''.
Where he argued that it is an disadvantage that we now have many independent scientific efforts
striving for the same solutions (that used to be just considered AI),
as it results in an enormous redundancy and hinders synergy.
In order to avoid this splintering, he suggests that AI should get back to its roots,
and again be an IPsI science (informational, psychological and intellectual), that covers all of the subfields:
A similarly fundamental importance as Physics or Biology. To
have a name for talking about it, we used ''IPsI-science'' (or
''Intellectics'' suggested by the author more than three decades
ago). While Physics is concerned with matter and Biology
with life, IPsI-science deals with IPsI-stuff (ie. informational,
psychological and intellectual stuff).
Of course ...!
3.5. Angry Birds - Competition.
According to the dictionary: ''Bird'', modern slang meaning ''young woman'' is from 1915, and probably arose independently of ''burd'', c.1300,
meaning, maiden, young girl.
But, surely, the organizers didn't have this in mind when they assigned rooms to these two events...?
According to the organizers at ''AI Birds.org'' (Angry Birds AI Competition):
The long term goal is to build an intelligent Angry Birds playing agent that can play new levels better than the best human players. This is a very difficult problem as it requires agents to predict the outcome of physical actions without having complete knowledge of the world, and then to select a good action out of infinitely many possible actions. This is an essential capability of future AI systems that interact with the physical world. The Angry Birds AI competition provides a simplified and controlled environment for developing and testing these capabilities.
In a traditional markets, people are paid to produce valuable resources.
Resources are sold at an appropriately high price, guaranteeing that the buyers had high value for them.
However, in many settings there might be alternatives to money.
Social media might give ''fame'', as ''payment''. We might want to
sell certain positions (Say, entrance to a certain school), not to those who are willing to
pay the most, but to those who are willing to risk the most.
Interestingly, whatever algorithm we will end up using (within a certain public
domain), it is problably important that the algorithm we use can be
''explained to a 5th grader'', if we want public trust.
Still, sounds logical, but it will of course still be interesting to see how many domains
of our lives will eventually end up being money free. Well, we will see...
4.2. Intelligible Intelligence & Beneficial Intelligence, by Max Tegmark.
Talking about the future of AI, 3 things becomes interesting:
Power. Artificial General Intelligence (AGI) is described as a point where AI's can match human intelligence in every category
- Is that a stage we will ever reach?
Steering. Do we get complacent, and allow machines to handle everything (in the future), while we do nothing,
and do not care where we end up?
Destination. If we believe that we will eventually see AGI's, should
these AGI's then be fully controlled by humans (disconnected from the internet) ?
Or should the AGI's take over (with our best interests in mind...) ?
Starting on the ''steering'' part, it is clearly not easy to trust AI's that we don't understand.
So, the more we understand how AI's work, the more likely it becomes, that we will be able to steer it,
at least a little bit.
So, we probably need to understand what is going in e.g. the deep neural nets (that are part of many AI programs):
Deep neural networks are now better than humans at tasks such as face recognition and object recognition.
They've mastered the ancient game of Go.
Actually, a bit puzzling, as there are orders of magnitude more mathematical functions than possible networks
to approximate them. And yet deep neural networks somehow get the right answer...
According to Tegmark, the answer is that the Universe is governed by a tiny subset of all possible functions/laws.
When written down mathematically, functions that are rather simple.
Meaning that deep neural networks don't have to approximate any possible mathematical function,
only a tiny subset of them.
In the talk, Tegmark didn't go that far, but maybe we do indeed live in a world, where algorithms rule, and where understanding
comes in the forms of other algorithms? Perhaps, even understandable algorithms... For more, see ''Solomonoff Inductions''.
Nevertheless (whether or not we live in a world of algorithms), today we still don't have understandable/explainable AI, and the public's trust
in AI is still rather hesistant...
So, what does this tell us about our (humanity's) destination? And what can we do to improve the situation?
Tegmark mentioned various ''AI safety'' initiatives that will help ensure that artificial intelligence will remain safe and beneficial.
I.e. the ''AI Safety Research program'',
funded by Elon Musk, has given over $2 million in funding to various
See more here.
4.3. Natural Language Processing and Computer Vision.
4.3.1. Rui Yan, Institute of Computer Science and Technology, Peking University.
Come to think of it, often, we probably don't even fully understand
all of the conversations we take part in... I.e. can we, always, precisely explain why we respond
the way we do..? Probably not...
A good conversation is all about exhanging information. We want to learn something
when we speak to someone.
And, all of this is obviously not easy to integrate into an AI system:
To build a conversational system with moderate
intelligence is challenging, and requires abundant
dialogue data and interdisciplinary techniques...
Generally speaking, we have two types of conversational
AI. One is focussed on helping the user with a particular task. Buying a ticket,
reserving a seat etc. Other chatbots deal with conversations in (more) open domains,
where the hope is that they will entertain us or give us emotional companionship.
Ideally, the intelligent conversational AI should be able
to output grammatical, coherent responses that are diverse
Which is a difficult task, even for humans...
Most standard chatbot systems today presume that only humans will
take the initiative in conversations, meaning that computers only
need to ''respond'' to the best of its capability. Giving the
chatbot a ''passive'' role. But this is clearly not how it works in human-human
conversations, where both participants can take the initiative.
So, well, there are still many problems to solve in this exciting area, before
we have chatbots that can pass the Turing Test.
4.3.2. Jing Lu et al., Human Language Technology Research Institute, University of Texas at Dallas.
Nelson Mandela had died aged 95 in Johannesburg. The world has lost a great man.
It might be easy for us to see that these two sentences, and events, are related.
But clearly not so easy for an AI program to see:
Event coreference resolution consists of grouping together the text expressions that refer
to real-world events (also called event mentions) into a set of clusters such that all the
mentions from the same cluster correspond to a unique event.
The tricky thing is that:
Event mentions are more diverse syntactic objects, including not only noun phrases,
but also verb phrases, sentences and arguably whole paragraphs...
Making this a very difficult problem. But, rather amazingly, some results have been achieved.
Hopefully, with more to come in the future.
4.4. Panel: The AI Strategy for Europe.
Cécile Huet from the Robotics Sector, European Commission, talked about the
AI strategy for Europe.
Why we should fund robotics research & innovation, a couple of headlines:
Essential for productivity and competitiveness.
Reindustrialisation, ageing workforce.
Advanced robotics is one of the key drivers of digital innovation.
An excellent talk, giving us some hope that Europe is still hanging in there!
Matuszek started by telling us that words are like icebergs.
Physically embodied agents offer a way to learn to understand natural language
in the context of the world to which it refers.
Human language does not exist in isolation; it is learned, understood,
and applied in the physical world in which people exist.
Connecting symbols (linguistic tokens and learning) to real-world percepts and
actions is at the core of giving meaning to these symbols.
Clearly, not an easy task...
Language-using robots must learn how words are grounded
in the noisy, perceptual world in which a robot operates... Still,
natural language systems can benefit from the rich contextual
information provided by sensor data about the world.
A super interesting talk!
5.1.2. Anca Dragan, University of Berkeley, USA.
Talked about Optimizing Robot Action for and around people.
Following the order that a human gives -- seems like a good property for a robot to have. But, we humans are not perfect and we may give orders that are not best aligned to our preferences. We show that when a human is not perfectly rational then a robot that tries to infer and act according to the human's underlying preferences can always perform better.
Self-driving cars that always ''play-nice'' might never be able to switch lanes on
So, sure, self-driving car shouldn't learn to be rude, but they should be able
to ''influence'' other drivers by their behaviour (i.e. ''make room for me'').
And, autonomous cars also need to be able to make good compromises between ''comfort'' vs. ''efficiency''.
Clearly, autonomous cars still have a lot to learn...
5.2. On Machines and Humans.
5.2.1. Spyridon Samothrakis, Institute of Analytics and Data Science, Essex, U.K.
Both the first industrial revolution, from 1760 up to
1840, and the second, from 1840 up to 1914, brought a significant change in working hours...
First, going from 180 days per year and 11 hours per day, to almost full year working and a 69-hour week
and after the first revolution.
Starting from 1870 data shows a consistent drop of working hours.
A trend that lasted until 1980 (when we observe a rise in working hours again, especially in the new world).
This tendency of decreasing work hours from 1870 led people,
like Keynes, to predict a four-hour working day (i.e. a 20-hour working
week). But that, of course, never materialized.
Society is being reshaped, and it is therefore obviously hard to predict what will actually happen
in the coming years. Still, people haven't stopped been interested in having ''The Good Life''.
And people certainly still want their luxeries - including Fame and Fortune.
will, of course, give ''money''- and ''social''- power. Even if you have to work hard to be part of that new industrial system.
Given our arguments above, we think the logical outcome
of improved automation is an increase in working hours, a
continuation of an existing trend.
5.3. Research Excellence Award 2017 - Andy Barto.
In this talk, Andrew Barto gave a brilliant introduction to some of the key ideas and algorithms of reinforcement learning.
Klopf recognized that essential aspects of adaptive behaviour were being lost
as learning researchers came to focus almost exclusely on supervised learning.
What was missing, according to Klopf, were the hedonic aspects of behaviour,
the drive to achieve some result from the environment, to control the environment
towards desired ends and away from undesired ends. This is the esential idea of
And reinforcement learning is therefore obviously (also) relevant in disciplines like machine learning, neuroscience, behaviorist psychology etc.:
Satisfaction to the animal will, other things being equal, be more firmly connected with the
situation, so that, when it recurs, they will be more likely to recur; those which are accompanied or
closely followed by discomfort to the animal will, other things being equal, have their connections
with that situation weakened.
Which brings us to ''learning with a critic'':
Where the ''critic'' returns the reinforcement, which evaluates the quality of the action (in the environment) taken by the agent
depending on its actual state.
5.4. Research Excellence Award 2018 - Jitendra Malik.
Malik has worked on many different topics in computer vision, computational modeling of human vision, computer graphics, the analysis of biological images etc.
And Jitendra Malik is also a part of Facebook's AI Research team.
When we recognize e.g. a bird, we do not only understand what shape
it has, we also have idea about what kind of texture the bird has.
In order to understand what action a ''thing'' is going to take, we need to understand
movement and goal.
I.e. Action = movement + goal.
Indeed, the human visual system is, of course, pretty smart.
A child doesn't need 1.000 examples of Zebra's in order to recognize a Zebra.
Children already have a lot (visual) knowledge that can be used in order
understand what a Zebra is.
So, artificial visual systems still have a long way to go.
Still, a very inspiring talk.
5.4. McCarthy & Computers and Thought Awards.
The conference ended with the McCarthy & Computers and Thought Awards.