The workshop was attended by 40 people, and sponsored by FIAS and our research project. We also a poster session with 10 posters, which I am not going to introduce, but you can check out the titles here. We had some remarkable keynote speeches, which I summarize below. More than that, we heard some important arguments and had a very insightful discussion about the science in "Constructive Developmental Science", "Developmental Robotics", "Autonomous Mental Development", or however you'd like to call it. Below I try to wrap up these arguments and try to situate them in context of recent debates in the ICDL community.
Our first speaker was Gregor from Ruhr-Uni Bochum as a representative of the dynamical systems perspective. He started right away with what that means: dynamical systems in the light of cognitive science. He said that, first of all, dynamical systems are a metaphor for cognitive systems. Not more. Not less. The "language" of dynamical systems allows us to think about certain things and grasp their properties. For instance memory or movement are things that are well viewed with the metaphor of dynamical systems. Other things are not. Like learning. Or rather, the underlying functioning of learning.
As a way to implement cognitive systems by utilizing this metaphor, we showed his lab's work on Dynamic Field Theory (DFT). A general introduction can be found in Gregor's last year's summer school slides. He show cased several introductory examples like sensory stimulus detection against noise, and memory traces thereof. An interesting application is certainly the modelling of Piagetian A-not-B errors.
Verena started by approaching the matter of intelligent from a sensorimotor perspective. Spoiling a bit of the discussion, we came about the famous quote:
"Why don't plants have brains? They don't have to move." [Lewis Wolpert]So, brains and intelligence have at least their root in motoric behavior, or more generally, as Verena put it with another quote:
Intelligence = goal-directed adaptive behavior [RJ Sternberg & W. Salter]Coming that way, she showed recent work with now PostDoc Guido Schillaci on flexible action selection and control by means of internal models. They trained independent forward and backward models for NAO's arms and them could select the right action (inverse model) by predicting success with the forward model. This is kind of MOSAIC-ish, but not exactly. Also she showed (submitted) work on self-other discrimination based on forward model's predictions. An interesting side news was that they also have experimental work on human sensorimotor learning now.
Kevin's talk was... different. Philosophical. So philosophical that it's even hard to correctly phrase the question he was trying to answer. The basic question her raised was
"How can organisms acquire a sense of space?"That said, it is important to say what he did not mean by that. I think (!) what he did not mean was that points in the visual field have a three-dimensional coordinate description. Or that there is some mathematical vector space in which things are described as a point. He was looking for an understanding of physical state. A space in which bodies are embedded instead of being singular points, and in which they move. A space that exists independently of the agent's existence. Honestly, I didn't see that coming. For me as a roboticist and machine learner a space is a vector space in which I project whatever.
here on arxiv.org. His first argument was that a sense of space could not be result of mere sensation. For instance, when things in the world's 3D space move, they cause a whole mess of changes of our retinal visual perception. Something that can not actually be described by three variables. Most importantly, how that change happens completely depends on the layout of sensors in our retina. It's not independent of the agent. The second argument copes with that: Suppose the world makes a "shift". It moves. When we are not mere sensation machines, but also act and move ourselves, we can do something interesting. We can compensate the world's shift, by shifting ourselves. We can look after an object after it has moved. And the extent you have to move is regardless of the sensori layout of your retina. So his argument was that when we can act ourselves, we can discover that the object and the agent are in one space in which changes can be compsensated. And because that is independent of the agent's sensors, that space also exists independently of the agent. Interesting argument.
Developmental robotics as well as developmental psychology mostly considers how babies develop, well, after they are born. Seems natural, no? Of course it does. But one should not forget they also the time a fetus spends in the whomb plays a vital role. It's not just growing in there. It learns, builds basic representation of the own body, and even listens to "outside" sounds. How vital all that is becomes clear when looking at preterm infants who miss some of that experience and fall behind.
So, how can we learn about and understand what is going on there? Tricky. One way is making "movies" of the fetus within the whomb by means of imaging techniques (that stuff totally blew my mind when I saw it for the first time). From a computational perspective, Mori pioneered working on fetal development and developed a "fetus simulator" on which learning algorithms can be run. Such as, you put simulated tactile sensors on it, process them with a neural processing model, and generate movement to trigger those stimuli again. Interesting finding: the model can, at least on a very qualitative space, explain which kinds of movements (e.g. hand towards the mouth) are observed how often. Only, however, if you take a human-like density distribution of tactile sensors, e.g. many on the mouth, very few on the back, and so on.
"All science is either physics or stamp collecting"With those words Rutherford once meant to criticize biological research for being purely descriptive, collecting just data on phenomena without giving coherent explanations. The same, Denis said (and I agree) holds for large parts of psychological research. We need computational views and models to put those things together and get an actual understanding.
As an example he showed a study about paradoxical results when infants learn categorizations. Suppose you teach infants what is a dog and what is a cat based on example pictures. Now you show them pictures of cats and dogs they haven't seen yet. Weird result: infants tend to think dogs are actually cats – but not the other way around. Why would that be? It can not really be explained by any hierarchy of concepts or categories. They tried the same thing just by training a neural network. And it does the same mistakes as the children, and the interesting thing is why. It turns out that often the statistical distributions of image features of dogs (thin line above) are largely included within the range in which cat-image-features (bold line above) typically occur. A classifier trained on such probabilities takes an actual dog, but when it falls into the overlap region it is misclassified as cat. And the amazing part of this story is, you can now change the distributions by manually picking examples. For instance such that the effect reverses, which gives a prediction to what might happen in an infant experiment if examples are picked accordingly. And the effect does reverse in infants. Cool.
Discussions and Thoughts
The interaction cycleOne of the discussions we had was about investigating the interaction cycle between the learning agent and its surrounding. One of the keynote speakers rightly mentioned the extremely data-centered view on learning that many researchers even in the developmental domain have. (Machine) Learning is the extraction of regularities from data. Where that data comes from? People often take it for granted. How that learning affects the later performance? We can think about that after closing our learning toolbox. What at least in some cases is fine for pure ML research is - and it seemed that everybody agreed - not a good account on developmental phenomena.
|Rolf, Steil, Gienger, 2011|
Too complicated.I have made somehow similar experiences. I have seen people referring to my work on Goal Babbling alongside with argumentations like "works on difficult problems, but it requires an interaction cycle". But? The more appropriate version – in my view – would be "works on difficult problems because it utilizes an interaction cycle".
One clearly has to say: yes, indeed, these interaction cycles are complicated. What happens in these loops is harder to understand than what happens "open loop". It's harder to implement properly. But it is crucial. Crucial for development (and also crucial for many "pure" ML problems). As discussions continued in the evening we all agreed on both the importance and difficulty. We need to strengthen our effort to investigate learning in the loop. And we need to take the experimental complication. What I think we could learn as well: it seems we have a communication problem. We need to explain much better and much more often why the closed loop is important.
Making testable predictionsAn interesting piece of discussion, and an important argument was raised by Denis. He made a very simple point. If we develop models for infant development, these models should – of course – make testable predictions. It's a strikingly simple point. But studies which actually do this are extremely rare to say the least. And that has a reason. Most people trying to come up with such models are engineers. If not in their education, then in their mindset. That is simply because robotics and machine learning require this mindset.
"A scientist describes what is; An engineer creates what never was." [Theodore von Karman]Engineers don't make predictions in the first place, they solve problems. Now what we do in developmental robotics, is trying to describe what is by means of creating what never was – at least when one sticks to that quote. Sounds a little paradox, but it is possible (see Denis' work). In fact, though, it really very rarely happens. I am no exception.
But if we (engineers) want to do developmental science, we need to stick to the scientific cycle of hypothesis, prediction, and experiment to finally come up with good theories. Yepp, it seems those loops are also there and equally challenging on a meta-level (imagine "science can solve difficult problems but it requires an interaction loop"...).
After all, this point goes in line with previous discussions at ICDL 2013 (see my post) of closing the loop between psychologists/biologists and roboticists. Giulio Sandini argued that, in order to tighten this loop, we need to move forward from descriptive models (explaining "how") towards explanatory models (explaining "why"). It is a similar argument, but not exactly the same. After all the focus on predictions as output makes a more concrete point on what to do. Plus, which is really important, it implies a way to organize actual work between actual people from different disciplines by engineers feeding those predictions back to experimentalists.
After all it was a pretty interesting workshop. The generally positive news was that most people showed their interest in continuing this line of "computational" workshops at psychologists' conferences. I am very much looking forward to that.
A further positive news to me and the community was that many people said they had submitted their work to ICDL-EpiRob. Let's meet there again!