Tuesday 19 April 2016

Brains learn to perceive, not predict, the world

The current trendy way to think about brains is as predictive machines. Bayesian methods, free energy, generative models, and all the other things in Andy Clark's new book are where it's at. Brains have to predict because, in Clark's terms, we are 'surfing uncertainty' - one step from disaster as we control our behaviour using sensory information we had to spend time processing and that is therefore a few hundred milliseconds out of date. Our brains build models of the world that constantly generate active predictions about what sensation will tell us next, and our behaviour holds together to the extent that the models do.

The ecological approach provides a different job description for the brain (Charles, Golonka & Wilson, 2014). We are embedded in environments rich in information that specifies a wide variety of behaviourally relevant properties. Some of that information is prospective, that is, it is information available now that is about something in the future. Two examples are the information variables you can use to catch a fly ball; both are states of the current time that, when used, allow you to achieve a behaviour in the future (specifically, to be in the right place in the right time to catch that ball). Another example is tau and the various other variables for time-to-collision.

This post reviews a paper (van der Meer, Svantesson & van der Weel, 2012) that measured visual evoked potentials in infants longitudinally at two ages, using stimuli that 'loomed' (i.e. looked like they were going to collide with the infants). The data show that the infant brains were not learning to predict the world. Instead, neural activity became more tightly coupled to information about the time-to-collision. We learn to perceive, not predict, the world.

van der Meer et al (2012) conducted a longitudinal EEG study on infants viewing looming virtual objects. The set up had the infants (tested at 5-6 months and again at 12-13 months) view visual stimuli that specified an object approaching all the way to collision under three different accelerations. The authors identified when the stimuli produced visual evoked potentials (VEPs), activity related to the stimuli, and assessed when the VEP peak occurred relative to collision time and the duration of that peak.
Figure 1. Babies in high density EEG rigs watching virtual objects fly at their faces, because science.
There were two differences in these dependent variables with age. The first difference was that the peak VEP occurred closer to actual collision time in the older children. The second difference was that the VEP duration decreased in the older children. In other words, the neural activity evoked by the looming stimulus became more accurate and less variable with respect to the timing of the collision. (This also makes the point that understanding what the brain us doing requires knowing what information it is interacting with - interpreting neural data requires a theory, and changing the theory changes the interpretation.)

A looming stimulus does not just create information about time to contact. It also moves with some speed and it changes apparent size (it gets bigger, i.e. it looms!). In a separate analysis, the authors identified when you would expect VEPs if the infants were responding to visual velocity information, visual angle information or actual time-to-contact information. At 5-6 months, 4 infants were responding to velocity and 6 were using time. By 12-13 months, 9 out of the 10 were responding to to time, i.e. they had shifted to the better variable. (The infant who was using velocity at 12 months actually switched from using time when he was younger; the authors note this infant was a late roller and had several months less crawling experience than the others). This data connects to the literature on the use of non-specifying variables.

To summarise: infant brains were not learning to better predict the world. They were becoming more attuned to information about time-to-collision, and the neural activity was getting better at preserving the temporal structure of that information*. We learn to perceive, not predict, the world.

References
Charles, E. P., Wilson, A. D., & Golonka, S. (2014). The most important thing neuropragmatism can do: Providing an alternative to ‘cognitive’ neuroscience. In Pragmatist Neurophilosophy: American Philosophy and the Brain, (eds. Shook, J., & Solymosi, T). Bloomsbury Academic.

Van Der Meer, A. L., Svantesson, M., & Van Der Weel, F. R. (2012). Longitudinal study of looming in infants with high-density EEG. Developmental neuroscience, 34(6), 488-501.

*If you think that sounds like a neural representation, you're going to laugh your ass off at our next paper, currently under review :) Stay tuned!

16 comments:

  1. It seems strange that you fail to note that the entirety of Surfing Uncertainty seeks to reconcile predictive processing with embodied/ecological/enactive approaches to cognition. Clark's argument is that prediction and perception are inseparable and both serve action in terms of affordances (ie, the central imperative of ecological cognition). You seem to be ignoring his argument for the sake of reinstating the very oppositions he shows to be false.

    ReplyDelete
  2. Since it is pretty clear you have not actually read the book in question, but have resorted to merely restating dogmatic oppostions, you should know that it includes an extended discussion of the classic ecological analysis of tracking a flyball, and takes that as one if its fundamental premises.

    ReplyDelete
    Replies
    1. I have a similar sense and at the same time hope that I'm wrong, as I wouldn't want to lose my trust in the rigor and honesty I have attributed to the writings here. I am curious about Andrew's reply.

      Delete
    2. I haven't read Surfing Uncertainty yet. But I've been reading and listening to Clark for years, and I know what he thinks the ecological/embodied approach should be like. He doesn't really get Gibson, or rather, he doesn't use the ecological stuff appropriately in his theories.

      In essence, while Clark uses many of the same terms, I simply disagree with how he uses them; I think he is incorrect. Using the same words is only part of saying the same things!

      My read on Clark's work is that he has not fully come to terms with the implications of Gibson. Not because he's dumb, he just disagrees with a bunch of it. But because he hasn't gone all the way, I find his use of the bits he does like to be weak and poorly constrained. His overall approach is strongly rooted in these predictive, generative models, and my stance here is that these things solve a problem that doesn't exist.

      Delete
    3. Thanks for your reply, that makes sense.

      As of understanding of Gibson's work, there's an interesting work from Stephen E. Robbins, an author who in my view understands Gibson and his implications thoroughly and ventures to extend them to quite challenging and non-mainstream conceptions (for example: http://www.stephenerobbins.com/uploads/7/3/2/9/73295531/nova-direct_memory.pdf)

      I'd be really curious on your view on his work...

      Delete
  3. If predictive models were indeed a viable means for surviving, then why do predators frequently fail to catch their prey? They are constrained by their senses to the same information that humans are. However, we have a better cognitive means for association and can apply our memories to broader contexts and make inferences.

    ReplyDelete
  4. What looks like a summary of the book in question: Andy Clark on "Predicting Peace: An end to the representation wars?"
    https://www.youtube.com/watch?v=d6EawXopUC4

    ReplyDelete
    Replies
    1. Yeah, I watched that; it's a good representation of his views (no pun intended :)

      Delete
  5. I just wanted to point out that there are radical embodied/ecological people who think that the theory of predictive coding/active inference can be made consistent with the central tenets of ecological psychology. Andy Clark is aware of this work and engages with it here, see especially his comments following the main blog-post: http://philosophyofbrains.com/2015/12/15/conservative-versus-radical-predictive-processing.aspx#more-4713

    ReplyDelete
  6. Excellent post. It is certainly important to understand the role an infants brain plays in assimilating information. And I agree, we can't treat brains as predictive machines. Although the brain simulates and predicts, these aspects aren't devoid of perception. As you've rightly pointed out the ball doesn't just create the information for 'time to collision'.

    I am particularly interested in the determinants of learning. And variation in stimuli gets highlighted so often. Without that variation, it would be harder for the brain to learn and predict in the first place. This variation in stimuli is largely about perception.

    ReplyDelete
  7. "The data show that the infant brains were not learning to predict the world. Instead, neural activity became more tightly coupled to information about the time-to-collision. We learn to perceive, not predict, the world."

    I'm trying and failing to understand how information about time-to-collision isn't a prediction about the world. Isn't it a prediction?

    ReplyDelete
    Replies
    1. No - it's currently available information that just so happens to allow you to coordinate your actions with respect to a future event. This is prospective control.

      The best way to think about this is to know that if you could 'read off' the state of the baby, nothing in there would correspond to the future event. It would all be about the currently available information. So the system doesn't know about the future in any way, it's just interacting now with information that happens to help.

      Delete
  8. "The best way to think about this is to know that if you could 'read off' the state of the baby, nothing in there would correspond to the future event. It would all be about the currently available information. So the system doesn't know about the future in any way, it's just interacting now with information that happens to help."

    So you don't think the time-to-impact "data" is being generated by a comparison with previous "states" and has any implication for how "future states" are generated?

    I guess I'm having trouble understanding how prospective control doesn't "incorporate" predictive control. I can see how they are different based on a few abstracts I've been poking around through but I still fail to see how the former doesn't include the latter in some way.

    Any good essays on that?

    I'm a really excited layman who has just discovered this REC stuff and am trying to get to a place where I can explain this to my friends.

    ReplyDelete
    Replies
    1. I try to explain prospective control here.

      And no, time-to-contact is not being generated via a comparison. The optical variable itself simply IS time-to-contact.

      Delete
  9. Perceptual Control Theory (PCT) provides a model and understanding of neural processes based on perception rather than prediction that explains how we function and survive in our changing environment.

    ReplyDelete