Information contents

“Information contents” is intended to be a new ‘topic’ IN THE EMBCOG postings.  The writing icon at the top right of PAICS has been clicked after opening the EmbCog category on the right (which gives past posts only). Ron, is this right? It’s the only way I can find to put a new title into EmCog itself. — David

[David, all you have to do to make a new EmbCog post is click on the writing icon and then after writing your post, remember to tick “EmbCog” under the “Categories and Tags” menu (and untick “Unclassified”) to the left of the edit screen.  I have done this for you now, so this post now appears under EmbCog — Ron]

Many thanks, Ron.  I’ll try both of those clicks next time. – David

Hi, Simon

Your paper on information theory for the Embodied Cognition reading group is excellent. What about putting it on Wiki Psychology?

I’m puzzled though why “Chemero’s discussion of information in chapter 6, and dismissal of Shannon information theory, irritated [you] so much” that you wrote this paper for the group. Chemero indeed wrote: “Information for perception is not Shannon-Weaver information” (p109) but this was in the context of discussing the contents of information delivered by affordances to the organism. At the end of your paper, you yourself emphasise that Shannon’s reductions in entropy ([tera]bits) do not deliver information content.

I wonder if the irritant is Chemero’s invocation of dynamic systems without pointing out that it is Shannon info which is generally used in that approach, whereby he contradicts page 109. I’m finding Chemero’s treatment increasingly irritating myself, as he keeps missing most of the point of what he refers to and even resorts to mere name-dropping. The most blatant is his glowing praise for neuroscience without specifying what’s so good about it, as a response to the charge that he is a “no-brainer” because he shows perception is performance of the animal while acting on its environment and so can’t be solely a brain (and body) process. The reduction of the animal’s long-term memory to long-term potentiation of some synapses does nothing to account for the content of memories (as I pointed out from Sussex in Psychological Bulletin 1967!).

This case against eliminative physicalism (of brain, with internal (bodily) and external environments – me again in Brain & Behavioral Sciences 1978) is what’s most important about the book. (Representation by bits of the system is a side-issue about more or less wise uses of that word, for Chemero as well as for you and me.) Yet Chemero makes a mess of key points in his Situated (not Embodied) Cognition.

Chemero’s section on the social environment is crucial because his detailed examples are about physical processes such as optics, skin contact, muscle contractions, stairs and blocks of varied shapes. Unfortunately, he does not seem to understand the hard causal structure of information in society: he even writes that affordances in society are “merely [sic!!] conventionally determined”. The ecological biologists are becoming very clear how powerful the traditions of successful activities are in primate groups. Physics itself (for Kuhn or Lakatos) is an extremely powerful and precise set of ‘conventions’ for effective handling of any environment (in its material aspects).

All the stuff about ecological optics, haptic touch and supposed extensions to other “energies” (a most eccentric idea for taste, smell, and hormone receptors in the brain) is far too close to Gibson’s outdated biophysics, materials science and kinaesthetic neurophysiology (and molecular genomics). Chemero’s only source of general theory of the information contents of physical affordances, let alone of social affordances, is “situation semantics” (unexplained). That buries in the deeper oceans of linguistics all the real and sometimes severe scientific problems of giving predictive accounts of the patterns that can be extracted from the stuffs and objects which are the sources of stimulation to sensory receptor arrays.

There’s quite a list also of conceptual moves that seem to me to be based on verbal muddles. E.g., contrary to misleading terminology in foundationalist epistemology and medical physiology, light does not cause sensations, except for dazzle (Hamlyn 1956). E.g., there’s no puzzle whether ‘perception’ (a misused noun) is inside or outside: the perceivING (the verb, active tense) is by the agent and the perceivED (same verb, in the passive) is in the environment.


– David



2 thoughts on “Information contents

  1. From: Simon McGregor McGill []
    Sent: 11 March 2016 12:21
    To: Embodied Cognition Reading Group

    Hi all,

    “Simon’s paper is considerably better developed than a larva! – at least a cabbage white butterfly, if not a red admiral… He should put it on Wiki Psychology?” (wrote David).

    Well… it was larval in the sense that I didn’t get as far as explaining why Chemero’s chapter is so annoying; I just laid the groundwork.
    I’ve expanded a little on that groundwork in my latest revision, but I’ll lay it out in more detail here.

    “I’m puzzled though why Simon wrote in his covering email, “Chemero’s discussion of information in chapter 6, and dismissal of Shannon information theory, irritated me so much …” Chemero wrote: “Information for perception is not Shannon-Weaver information” (p109) in his discussion of the contents of info delivered by affordances to the organism. At the end of his paper, Simon himself emphasises that reductions in entropy / [tera] bits do not deliver information content.”

    I absolutely did not say that! What I said was that the quantity I(X;Y) does not represent the content of information in X. It obviously can’t, because it measures the amount of something. What is that something? In my opinion, it is precisely the richness of information content. Using a loose analogy, mass does not represent the contents of a sack, but it does provide a meaningful measure of the amount of stuff in the sack. And if you want to understand stuff properly, you won’t get far by insisting it has nothing to do with mass.

    Chemero did indeed write “Information for perception is not Shannon-Weaver information” (p107), but he immediately followed it with an example that totally undermines his claim:

    “In the fog-filled room, the light that converges on… an observer’s head and eyes has been scattered by the fog… In the more typical, nonfoggy situation, the light that reaches any point in the room has been reflected off the room’s surfaces… This set of facts is what allows the light that converges at any point to carry information about the substances in the environment.” (p107)

    My paper explains the sense in which this information can straightforwardly be measured in Shannon bits: a hypothetical Bayesian observer, who does not know the configuration and composition of the surfaces, will become more subjectively confident about those things after knowing what light converged on their eyes in the clear room, but not in the foggy room. In other words, in the foggy room the light literally carries no (Shannon) information, whereas in the clear room it does.

    This information-theoretic perspective also provides – for free – insights that Chemero lacks. Information is relative to a subjective ensemble; suppose (for whatever reason) we choose to consider a range of environments in which the the configuration of surfaces can (at least in principle) be deduced from the light striking the eyes in a foggy room, due to severe restrictions on the considered microstates of the agent, fog and surfaces. In this range of environments, the light that converges on the observer’s head and eyes has been scattered in a way that is highly predictable. Despite this, the scattered light will not cause the agent to behave in ways that are appropriate to the surface configuration (unless we make our initial ensemble even more pathological); we may say that (in such a range of environments) there is information about the surfaces even in the foggy room, but that the agent cannot use it.

    Of course, we do not really care about the systematic behaviour of that ensemble; it seems highly implausible that it will ever help us to understand the behaviour of agents in the general, more restricted world. When we talk about “the information” without qualification, we are implying the proviso “in any scientifically interesting ensemble”.

    Much of what Chemero goes on to say about information (which, remember, he claims is not the information measured in Shannon bits), can be said more precisely and succinctly in statistical information theory. I’ll provide a few illustrative examples below.

    “In particular, the light converging on some point of observation is in a particular relationship to the surfaces in the room, that of having bounced off those surfaces and passed through a relatively transparent medium before arriving at the point. The information in the light just is this relation between the light and the environment.”(p108)

    Yes, and I have tried to explain how its units of measurement are indeed Shannon bits (contra Chemero and Gibson).

    “Because Turvey, Shaw, and Mace take direct perception to be infallible, they insist that it be underwritten by information, which is, in turn, underwritten by natural law. They are careful to maintain that the laws in question are ecological laws, laws that hold only in particular niches.” (p112)

    Again, this relationship of information to ecological niches can be derived a priori from considering Shannon information over subjective ensembles.

    “According to [Barwise & Perry’s] situation semantics, information exists in situations, which are roughly local, incomplete possible worlds.” (p116)

    Everything Chemero excitedly derives from this idea, he could have got for free (and with far more generality and precision) from statistical information theory. In the expression I(X;Y), the X and Y refer to “random variables”: these refer to particular features of a system when considered in isolation from everything else in the system, i.e. “local, incomplete possible worlds”.

    “Because constraints need only be reliable and not lawlike,
    nonspecifying variables can carry information. Millikan also makes a valuable point concerning just how reliable nonspecifying variables need be. On her view, the correlation between two events need be just reliable enough that some animal can use it to guide its behavior.” (p119)

    Comes for free from statistical information theory. But statistical information theory can go even further: we can ask quantitative questions about the causal flow of (Shannon) information from the external variable to the animal’s behaviour through the channel of sensory stimulation. Since the picture is not simple for embodied agents, the answers will be complicated and probably instructive.

    In almost the whole of Chapter 7, whenever Chemero refers to information about affordances, the word “information” can be interpreted directly as a thing whose richness is measured in Shannon bits (in the manner I describe in my article). The one exception is as follows:

    “[Dynamic touch] information is correctly formatted for action.” (p159)

    Statistical information theory has nothing to say about the “formatting” of information (and nor, for that matter, does Chemero). It seems to me that the notion of formatting is very poorly-defined, and that there can be no sense in which information is “correctly formatted for action” that does not make implicit assumptions about the agent’s internal mechanisms. Rather than talk about how information is formatted, I am more comfortable talking about how the agent’s sensors, actuators and control machinery combine to transmit available information into actions.

    Warm regards,


  2. From: David Booth Sent: 11 March 2016 15:38 To: EmbCogRG

    Hi, All

    Simon is correct (of course) about the many things that Shannon information can do that are (or should be) of interest to Tony Chemero. However, as far as I can see, the Chemero book we’re reading says nothing about the actual substantive contents of the information in the amounts measured by Shannon. Simon’s reply continues to make clear that he does not say anything either about the contents as such. Simon writes (today):

    “What I said was that the quantity I(X;Y) does not represent the content of information in X. It obviously can’t, because it measures the amount of something. What is that something? In my opinion, it is precisely the richness of information content.”

    In short, I(X;Y) measures richness of information in X. Simon illustrates this with Chemero’s examples of lots and lots of information versus zero or very little information.

    As Shannon said, that measure of amount (great or small) conveys nothing of the content of the transmission, or of the transmission’s effect on the content of the receiver’s state, e.g. specifically what is perceived, remembered, thought about, felt or acted on.

    For example, the content of visual information from an approaching vehicle to the driver on a clear or foggy night includes the difference in number of lumens (or related units of physics) transmitted from a headlight and (say) reflected from cats-eyes down the middle of the road near the headlight, within a wider visual field. There is zero difference in luminances (i.e., it’s invisible) when the fog is thick enough, but on a clear night there is a high contrast to be received, alongside many other distances between locations of specifiable luminance values – as also for the rat navigating its tank.

    In other words, “I(X;Y) measures the amount of [information] in X”; “it can’t” … “represent the content of information in X.” “Content” refers to the actual physical (and societal) quantities, categories and patterns that are (momentarily) conveyed by whatever richness or poverty of info. the agent is using.

    – David

    P.S. Simon and I agree about Shannon reduction of Chemero’s “formatting” too. I take Chemero’s word “format” to be his attempt to refer to the structure of the content of information. Structure is an aspect of any concretely specified information content, e.g. the way in which two or more physical and/or societal parameters are ‘combined’. This inclusive and therefore vague term ‘combining’ encompasses a specified variety of determinate mathematical operations on environmental data, in all sorts of units which have each been transformed into a universal unit of causal strength with respect to their influence on action. – DAB

Leave a Reply

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

You are commenting using your 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