ECAL 2005
2 weeks ago I was in Canterbury for ECAL 2005. I've taken a while to write about it because I wasn't quite sure how to structure a conference review. I'm still not, but behind the cut I'll rehash the report I sent to the rest of my lab so as to at least say something about it.
First of all some online resources:
- The workshop proceedings (not included in the printed proceedings volume) are online
- Some talks were recorded, and some speakers interviewed for Complexity Digest. Full audio of certain talks, and video summaries of others are available at the ComDig website
As far as I know there's no central repository of the papers online, but of course some authors self-publish on their own websites.
[I give a lot of partial references below. Unless I specify workshop paper
, they are all things contained in the main proceedings volume]
Overall impression
I think this conference had the lowest ratio I've experienced of papers that made me think well that's nice but so what
- things seemed to be quite well-focussed and with a point.
There was, as expected, a strong Sussex contingent. What I learned from them is that there's more work than I had realised over there that picks up directly from our lab's projects, so that will be worth watching over the next couple of years. In particular, Eduardo Izquierdo-Torres has research interests that match up very closely with mine and there may well be several Sussex graduates applying for postdocs at Indiana University.
I think I'll try to organise the rest of my general comments around subject areas that I thought were both well-represented at the conference and of interest.
Learning and memory
There are several people looking into problems relating to learning and memory, and indeed one of the workshops was on this subject. I got a few general messages from this part of the conference. One is simply that the evolution of learning is a problem of interest to people, but more importantly the conceptual issues I've been struggling with of how to classify different types of learning are regarded by others as worthwhile open problems. There were a few presentations that addressed the difference between simply learning to choose between two behaviours and actually being able to cope with novelty, though I'm not sure the actual experiments I heard about make the distinction quite clear enough. Trouble is, I'm not sure mine do either.... [the Memory & Learning Mechanisms in Autonomous Robots workshop was where these things were discussed the most]
Other things that were talked about on this subject included confusion between learning, evolution and what exactly the Baldwin Effect means [the Mills & Watson paper in the main proceedings; I can't find it online], and the importance of forgetting [offline discussion, though apparently this had been inspired by a presentation I wasn't at]. I think I've been underestimating the importance of forgetting with my own work - the point is basically that it's all well and good making an agent that can learn one thing from a randomly or uniformly initialised starting condition, but it may be an altogether different challenge to make one that can continue to adapt as its environment changes. Of course we see both processes in biology, in the same organisms.
There was also an interesting conceptual paper by Eduardo Izquierdo-Torres & Ezequiel di Paolo (paper and slides) addressing the question of how we distinguish between a reactive agent and one that has internal dynamics; basically they suggest that the distinction is not clear, though I think I would define reactive
differently from them. I need to read this paper before saying anything more about it; it was certainly food for though.
Theory of evolutionary computing
There were a lot of talks about theoretical issues in evolution. It was the track with the most papers, and I didn't feel that there was one theme I could summarise, so instead I'll point out a few particular papers:
Schlessinger, Bentley & Lotto [none of their websites have the paper up yet, but they all publish online so it will probably appear]: "Analysing the Evolvability of Neural Network Agents Through Structural Mutations". Basically they tested 5 different kinds of mutation operator, which caused different sized changes to the neural network, and compared the performance of evolution runs using each type. They found some differences, but I'm not sure that they are appropriately comparing like with like, because the amount of change in one generation is quite different between mutation operators. In any case I liked the idea of what they were doing, and I'll have to go through the paper properly to see if that objection is addressed.
Simon McGregor presented a paper about generalisation. Basically he found that evolved digital logic circuits can cope with only seeing part of the logic table they're supposed to be instantiating, provided that they see enough for there to be a correlation between the missing rows and the rows they were evolved on. I think this might have implications for generalisation in other sorts of evolved controllers.
Polani, Dauscher & Uthmann: [not on any of their websites yet] On a Quantitative Measure for Modularity Based on Information Theory
. As the title says: this is an attempt at putting together a precise measure of modularity, which they then apply to the selection of crossover operators.
Biology and psychology
There were several papers by people who considered themselves biologists or psychologists first, and others by alifers whose main aim was to explore specific biological phenomena. As a general point, it's nice to see some bona fide life sciences people taking modelling seriously as a tool, and specifically there was a dictyostelium paper that seems worth reading:
John Bryden: Slime Mould and the Transition to Multicellularity: The Role of the Macrocyst Stage
. It uses a very different sort of simulation from the amœba work in our lab, because the aims are also different, but I think it's still relevant. More than looking at the mechanisms of aggregation, this paper is focussed on looking at why the aggregation happens in terms of how there is a benefit to individual cells from joining the aggregation.
Swarm robotics
I remember there being quite a lot of fuss about swarm robotics at both the conferences I went to last year, and I remember not really understanding why. This time there was a plenary talk by Marcus Dorigo of the swarm-bots project, in which he presented considerably more impressive results than I had seen before. Basically my impression is that now they've gone beyond "hey look we can make some robots join together like Voltron but we can't actually get them to do anything" to actually getting them to do something useful once they have aggregated, and that shift to goal-directed behaviour seems significant. What I appreciate about this work is that it's all done with decentralised control and purely localised signalling; in some ways it seems conceptually tied to natural phenomena like ant colony organisation and dictyostelium aggregation.
Being a plenary talk there's no paper to go with it, but if you're interested the swarm-bots project has a homepage.
Implications for me
As well as the science, there were a few personal things I took from talking to people offline at the conference:
- There actually is interest from at least a couple of people in the projects that I've left behind for the time being; specifically in analysing the controller networks of the catcher agents, and in trying to get the landmark learning agent working properly. Right now I think I do need to focus on a research proposal, but once that's done I would like to pick the landmark learning project back up, because I think there's at least a paper in that without having to spend months on it (like I would to get anything substantial together with the analysis of the catcher agents).
- I need to get some work out there, because that does generate useful feedback.
- The ALife X deadline is closer than I thought: November 7th. I don't know if I'm going to have anything worth writing about by mid-October, but I should try. Failing that, the workshop deadlines will be much later, because the workshops themselves haven't been announced yet.
Comments
This makes me think of a few things... (and pardon if I royally go out-of-the-blue here, because most of what you do is probably out of my area).
1) The entire slime mold aggregation thing makes me think, of all things, of oral cavity biofilms - collections of bacteria in the mouth that form a protective sugar coating and hang out in the crevices in your teeth in a sort of mutualistic hippie commune. I wonder if anyone has started to look at aggregation of disparate microbial species?
2) Human brains are programmed to selectively forget "unimportant" data. How do you model that in a virtual system? How does a model "choose" what to remember and learn from? Is there a period of "sleep" in models that allows for integration of knowledge and "stream-of-consciousness" connection between new data and already stored information?
Alexis,
Sorry it's taken me so long to answer. I had a trip to Cleveland last week, before which I was scrambling to make sure I actually had enough to say to make the trip worth it, and after which I got sick. I have been meaning to reply to this comment for some time because what you ask is interesting.
1) I don't know of any such work myself, but I'm sure it does exist. In technical terms, modelling with multiple species is more difficult than modelling just one, but it also allows us to look at some other interesting questions, and symbiosis is in general a fascinating problem because it always poses the question of why organisms don't cheat.
Can you point me to any references about oral cavity biofilms?
2) You're asking several questions which are beyond the scope of the super-simplified models I'm playing with right now, but also worth studying. I'll respond to different parts individually.
Selective forgetting: my model doesn't really let me explore this because there's only one type of information in the simulation, so the only distinctions are between "current" and "out-of-date" information. What I am trying to build is a system that can discount the value of older information, so that instead of only learning about the world once and then being set (which is how must such simulations operate, but it's arguably better described as "development" than "learning"), it can continue to be flexible. So in my models, the "choice" should be made purely in terms of recency.
In biological systems, clearly much more than this goes on. I'm familiar with some of the psychological theory of learning and forgetting, and from memory it points to several different influences:
- Serial order: primacy and recency effects, i.e. the most recent information and the earliest hold the most sway.
- Levels of processing: that which has been processed more deeply is more likely to be remembered
- Emotional salience: if an event or piece of information is associated with anything else of emotional significance, it's much more likely to be remembered.
- Survival value: organisms seem much more likely to remember information that would be relevant to their survival in an ecologically realistic setting
- ...there are many more, but you get the picture
Clearly only #1 applies in a model where the agents only get one kind of information and only have one very simple kind of behaviour, but in a more complex system the others could certainly be built in."Choice" of memories: I think this really relates to #3 and #4 above, but I sidestep the question by only having one type of information/event (after eating, the agent's energy level is either increased or decreased) for the agent to attend to. Some of the experiments I'm running do have random inputs that provide the agent with no information whatsoever, but ignoring those is too trivial to be of much interest.
Sleep: Again this doesn't fit into my model because of its sheer simplicity, but is probably relevant for learning systems in general. I think another issue is that I'm totally ignoring synaptic plasticity—all information storage in this model is in the neurons' current states (and can be maintained by recurrent connections)—so there isn't the variety of learning process speeds that I suspect is one of the reasons sleep is important for real organisms. I suspect [this is very hand-wavy] that sleep has an important role in the gradual transfer of important information from short-term learning processes (such as the learning that exists in my model) to longer-term ones like the physical growth/pruning of new/redundant dendrites, and changes in the organisation of the brain.
Hi Eldan,
Just stumbled across your blog by googling for "Slime mould and the transition to multicellularity" and, after reading this blog entry and then searching for more information about the author, ... Hey, you were the guy who also stayed at Kipps, right? :)
Anyway, I just wanted to point out that there actually _is_ a central repository of the papers online. As the publisher was Springer, all the papers are available through http://www.springerlink.de/
However, the bad news is that unless your institution has a subscription you have to pay for every paper.
The conference was indeed quite informative and I got a lot of new ideas from there, too.
Best wishes,
Taivo,
That's right, we were staying in the same hostel. I'm glad you found me, because I was entirely unaware of the SpringerLink availability. I'm not sure whether my own institution is a Springer subscriber, but I'll link to it anyway for the benefit of anyone who is:
http://www.springerlink.de/link.asp?id=an4b7fv1umg8
It might be useful to someone, anyway.
In terms of references to biofilms, I don't have many - it's more of a concept that was discussed in a few classes than anything else. Here's one on wound biofilms that seems to be a good general guide. this might be more what you're looking for. It's not really something I know much about; I just think they're cool.
At some point in this response to your response, you're probably going to say "she doesn't understand computing" so I thought I'd just mention it now - I have no friggin' clue. that said:
Is there a way of assigning numerical relevancy to ideas in such a way that the numerical value changes over time - some sort of series equation, etc... ? Moreover, can you assign differently changing numerical values to different kinds of memories?
I'm sure this is beyond the scope of a simple model, but ideally, my thinking is saying:
Emotions could be assigned different diverging/converging numerical weights that would change over time, and begin at an assigned point in the change relative to how impactful the researcher wants the emotion to be. The numerical value will then determine the memory's weight in the "mind" of the model, beginning at the n=0,1,2,etc value of assigned weight and either expanding to more weight or decreasing to less over a period of time. The weight of the memory and it's particular inherent weight (negative reinforcement vs. positive) as assigned by the researcher could then somehow impact choice.....
This period of time could also be controlled so that changes in numerical value occur on a set schedule. Thus a "down-period" in which nothing new was being introduced and yet numbers were still calculating could be considered "sleep." Once a set of numbers reaches a certain value-level, they could assigned as "long-term memory" and constitute some sort of deep processing.
"Long-term memory" could in turn be given a separate, very slowly decreasing set of numerical values to deal with loss of memory over time....
Values could be set to interact with each other at specific points in the sequence randomly - memory A just happens to return the same value as memory B at time point Q, in order to simulate free association.....
Choice of memory could be somehow correlated to numerical weight at a given timepoint....
all somehow. This is all way beyond me, but it seems to me that there's got to be a mathematical way to respresent the decay of emotional value of a memory over time and thus its impact on choice, and moreover that there must be a mathematical way to represent long-term vs. short-term memory, and that if there is such a mathematical way, then a computer model should be very possible......
or am I completely off target?