The Life & Mind Seminar Network

Seminar: The Mind as a Temporal Resonator

Posted in Seminars by matthewegbert on February 13, 2012

This Wednesday, I will be presenting the ALERGIC/ Life and Mind Seminar at the University of Sussex.  I’ll be introducing a completely new model, and trying on some new ideas that I’m excited by.  I hope to see you there. All are welcome!

There have been some requests that I try to record this one, so if I manage, I’ll put a link to the video here.

Cheers,
Matthew

 

The Mind as a Temporal Resonator: A new model for the study of autonomy, cognition and repetitive behaviour
Matthew Egbert
Room Arundel 401 @ 4.30 pm, Wednesday 15th Feb

Abstract
I will present a (very) new computational model of a cognitive system.

At the heart of the model is the notion that we (cognitive systems) tend to do what we did before, when we were in a similar situation in the past.

When this property is included in a dynamical model of an embedded cognitive system, self-maintaining patterns of behaviour emerge that involve (and depend upon) a wide variety of complicated interactions with environmental features. This occurs in the absence of any system of reward or punishment.

The control system in my model is based on the idea of an “Imprintable Dynamical System”, a new dynamical construct that is a little bit like “hebbian learning in state space”, but different. I will explain this construct in the talk.

The research is inspired by Gordon Pask’s “Ear”, the notion of autonomy as discussed by Varela and Maturana, and work on habits undertaken by Barandiaran, Di Paolo and others. It takes the idea of autonomy discussed in the biochemical domain (autopoiesis) and demonstrates how these ideas can be applied in the cognitive domain, showing how a robot doesn’t need to be autopoietic to be autonomous.

The model outlines a way of thinking about cognition that may help us to better understand forms of mental illness, addiction, repetitive behaviour, adaptive behaviour, autonomy and cognition.

Advertisements

8 Responses

Subscribe to comments with RSS.

  1. Tom Froese said, on February 14, 2012 at 2:18 am

    Wow, sounds very interesting! I hope you’re submitting this to the upcoming Alife conference?

  2. marekmcgann said, on February 14, 2012 at 3:36 pm

    Sounds awesome indeed. Will the talk be podcast?

  3. matthewegbert said, on February 14, 2012 at 4:13 pm

    @Tom: Unfortunately, I don’t think I’m going to be able to make it to ALIFE this year…so probably not.

    @Marek — it should be recorded. I’ll post a link here when it goes online.

    Cheers,
    Matthew

  4. […] you want to discuss or comment on the talk,  this page of the Life and Mind Blog would be a good place to do […]

  5. Ezequiel said, on February 18, 2012 at 2:02 pm

    Great talk and model, Matthew. Would be nice to discuss the details further. Just a couple of thoughts.

    It reminds me of Pask’s ear, of course, but also a model of path formation on grass by Helbing (using simulated individual agents walking on a plastic medium – the grass – somehow the vector update in your space remind me of the update rule they used). Also some similarities with the Ashbyan model of spontaneous change in preference we’ve done with Hiro Iizuka.

    I think that some parameters of the update rule for the vector space can make all the difference to the results. I’m thinking in particular of the relative effect of the strength of the update rule and the decay factor (which it wasn’t so clear how it works – does each vector decay to a random value?). I would also suggest having a volumetric neighbourhood update rule (not just along the axis lines), using a Gaussian for the update strength. Actually, it would be interesting to see if the effect of using a combination of a positive and a negative Gaussian, so that you strengthen local values to be like the new updated value, but far-off values to be move in the opposite direction. This should bring some definition to the structuring of the whole space (something like this is used in several Hebbian like rules – local potentiation, long-range depletion).

    I think that the real test for the potential of such a system to model cognitive autonomy will be when you put it in situations where different ‘habits’ interact with each others (e.g., ambiguous sensorimotor situations that overlap between habits), and even more so, in situations where there is a change in environmental dynamics so as to revert the ‘value’ of a behaviour (e.g., inversion of vision, inversion of profitability to a certain action, and so on). In such cases, this model, since it doesn’t involve explicit reward strategies, should prove superior to various kinds of reinforcement learning because the value of a behaviour is what shifts and therefore what previously produced a reward could now be a bad choice. (And having metaplastic rules simply defers the problem one step without solving it.)

    I think that what I’m suggesting is to move this towards what you’ve shown with the metabolism-based phototaxis models. Once you cash out the notion of the ‘value’ of a behaviour in dynamical terms using these self-reinforcing sensorimotor patterns, then you have just one mechanism explaining behaviour, plastic adaptation, and metaplastic tracking of changing values.

  6. Nathaniel Virgo said, on February 19, 2012 at 8:12 pm

    Hey Matt

    I’ve just watched the videos. This is really great stuff. I’m sure there’s a lot of potential in this approach – there are very few systematic approaches to learning in a truly dynamical context (at least that I know of), so this could be quite important.

    An interesting idea to avoid the curse of high dimensionality might be to conceive your system not as an entire controller but as a node in a network. So the controller would be composed of a (possibly recurrent) network of many “neurons”, each connected to only a few other neurons, and each with its own low-dimensional phase space, composed of its inputs and output. Of course, this introduces the additional problem of needing to determine a good topology for the network – but still it might be interesting to try.

  7. fcummins said, on February 25, 2012 at 9:13 pm

    Very good talk, and a very elegant model. I think the simplicity of the core of the model makes it a good candidate for informing our discussion of the larger framing issues, such as autonomy, etc.

    Some relevant prior art not raised in the discussion is Jaeger and Eck’s model of the earworm phenomenon, which got an echo state model to enter a repetitive cycle based on environmental exposure. That model suffers from a rather baroque, and theoretically unmotivated, “voting system”, but the core of it is the induction of any of an almost infinite set of periodic attractors in the dynamic core of the ESN. (Scholar search for “Can’t get you out of my head”, Jaeger and Eck, 2008)

    The dynamic nature of the environment certainly does not look like a “cheat” to me, but I would regard it as an essential element in the model if we are to take it seriously as a model of living beings in real environments.

    I would like to understand the role of periodicity a bit better. Your dynamic environment is simply periodic. It seems with simple models, we continually constrain our focus on simple forms of repetition (the Kelso work on finger wiggling comes immediately to mind). As this work progresses, I think it will throw up very interesting questions about the patterns we interpret as being repetitive, and with that, the relation of “thingness” to our definition of time. I hope to take up the discussion with you some time some place!

  8. matthewegbert said, on March 3, 2012 at 3:30 pm

    Hello Everyone,

    Thanks for the comments and sorry for the delay in replying. Teaching has been demanding my time recently, but now that term is coming to a close, my obligations are fewer and I’m excited to have more time to think about and work on these ideas. There are many directions to take the model, and I’m still working to figure out which would be best. Any ideas about connections to experimental psychology, machine learning, or other areas would be much appreciated!

    Here are a few comments and replies to your comments.

    ** Update Rules **
    People here and elsewhere have commented / asked questions about on the update rule. In the talk I presented, the update rule was what I’ve been calling the “plus sign” update rule: lattice points that have a similar value of the current state in all-dimensions-but-one are updated so that they point more in the current direction of travel through state space. This update is scaled by the distance (in state space) such that the closer the dfield node, the stronger the update. As I think I mentioned in the talk, this was conceived as a way to combat the problem that the higher dimensional delta-field, the less likely it is that you will return to a point in the delta-field that you have been at before.

    There are many ways to modify the update rule (like the Gaussian method that Ezequiel mentioned). This is something I plan to work on, but initially, I’m most interested in keeping the update rule very local in the dspace. The “local-potentiation / long-range depletion” idea is intersting, but it makes it difficult for multiple habits (or more accurately, the delta-field patterns that enable them) to persist. One habit circling around would erase any other habits elsewhere in the delta-field. Multiple habits, and their interaction (as a few people have commented) are a very interesting idea to pursue in this model.

    Another idea for the update rule would be to include a tendency for proximal, but unvisited regions of state space to tend to move towards visited regions.

    ** Decay factor **
    If I recall correctly, the decay factor is pretty minimal in the videos I show. It was mainly included at all so that the idea of self-maintaining pattern could be more easily defended. But if the decay rate is too high, again, it’s impossible to have multiple habits. This is one of a few parameter that I’m not sure how to set..

    ** Self-sensitivity **
    If I understood your last comment correctly, Ezequiel, what you are suggesting I include is a way for these habits to be, in a sense, self-sensitive… for them to react or change somehow in a way that is in response to how well their self-maintenance is operating. This would be a very interesting addition, I have one half-baked idea about how to go about doing this. I hope to get to it this summer!

    ** Training dynamical systems **
    Thanks Nathaniel for your comment about there being few systematic approaches to learning / training dynamical systems. This is, perhaps, an application of the idea that might be useful in terms of grant applications etc. Many of us have used genetic algorithms to train CTRNN, but I’m curious if there are any other techniques that people know about for creating stateful continuous dynamical controllers…? If not, maybe that’s a good direction to take this research.

    ** Periodicity and rhythmic behaviour **
    Thank you too, Fred, for your comments, I am interested in exploring more complex environments. The periodicity of the environment, and the motion of the light around in circles are indeed why the phototaxis is one of the most stable patterns of activity. Figuring out what makes for stable patterns is one of the main things I’m curious about at this stage. I also think that rhythmic behaviour might be an area where this model could connect with experimentation. Any pointers to papers that might be relevant would be greatly appreciated!

    ** Interaction **
    This is the first simulation I’ve ever created where I’ve really wanted to be able to interact more directly with the agent. I’ve been using a playstation-type controller, but it still would be better to have a robot in reality that I could push around… One idea I have is to couple this controller to a microphone and speaker and to try to see what kind of feedback patterns are stable. This is the simplest way I can conceive of to seriously enrich the environment that the controller is in. But so quickly these systems become so very complicated!

    ** Current quandaries**
    Right now, I’m trying to figure out principled ways to decide upon
    some of the parameters that significantly influence the dynamics of the system..

    – how fast does the dfield decay?

    – how significantly does a given trajectory through the dfield modify the dfield — in other words, does it take 2 passes to “enscribe” the path, or 200?

    – how many discretizations in the field are there? How to include this in a way that the dynamics are (approximately the same) no matter how many discretizations there are.

    Hoping to have lots of time to explore these ideas in the near future.

    Cheers,
    Matthew


Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: