Leiter Reports: A Philosophy Blog

News and views about philosophy, the academic profession, academic freedom, intellectual culture, and other topics. The world’s most popular philosophy blog, since 2003.

  1. Fool's avatar
  2. Santa Monica's avatar
  3. Charles Bakker's avatar
  4. Matty Silverstein's avatar
  5. Jason's avatar
  6. Nathan Meyvis's avatar
  7. Stefan Sciaraffa's avatar

    The McMaster Department of Philosophy has now put together the following notice commemorating Barry: Barry Allen: A Philosophical Life Barry…

More on whether evolution really selects for veridical perception

Brian Leiter

Cognitive scientist Donald Hoffman (UC Irvine) kindly shared with me his response to the criticisms of philosopher Jonathan Cohen (UC San Diego) I noted in an earlier post about a popular article of Hoffman's at the LARB.  Professor Hoffman's response:

We replied in 2015 to Jonathan Cohen's claim that (1) we misunderstand representation, (2) we argue for veridical perception of fitness payoffs, and (3) we must specify the contents of perceptions before we can evaluate them for truth or falsity. The reply was published in Psychonomic Bulletin & Review: https://link.springer.com/article/10.3758/s13423-015-0931-3

Here is an overview of our reply. (1) Cohen states that our definition of veridical perception requires an absurd identity theory in which x veridically represents y iff x = y. We reply that this is a misunderstanding of our text that is easily set straight; we do not require the identity x = y. We simply require a homomorphism. (2) Cohen argues that it is obvious that we must specify the contents of perceptions before we can evaluate them for truth or falsity. We reply that it is obvious that we do not need this. In formal logic, we do not need to specify the contents of p and q before we can conclude that p ∨ q ⇒ p is a fallacy or that p ∧ (−p ∨ q) is false. In information theory, we do not need to specify the contents of messages before we can correct errors in their transmission. Similarly, in the logic of evolutionary games, we do not need to specify the contents of perceptions before we can evaluate them for truth or falsity. The power of logic, information theory, and evolutionary game theory is to arrive at truths that are independent of assignments of content. (3) Cohen proposes specific content assignments for the jewel beetle and dragonfly examples so that their perceptions turn out to be veridical, and then claims that we are not entitled to make content assignments on which they turn out to be nonveridical. We reply that our evolutionary games, which were formulated with no a priori assignment of contents to perceptions, do allow us to rule out the content assignments that Cohen proposes, and moreover to rule out all theories of perceptual content currently proposed by philosophers of perception. (4) Cohen argues that our evolutionary games show that perception veridically represents fitness payoffs. We reply that fitness payoffs, unlike states of the world, are not organism-independent features of the world: They cease to exist if the organism ceases to exist. Perceptual experiences are satisficing solutions to the problem of getting more fitness points than your competitors; they are not veridical representations of fitness functions or of objective features of the world.

I've opened comments for those who wish to engage Professor Hoffman's rejoinder or Professor Cohen's original criticisms (or the original article).   Comments are moderated; submit your comment only once, and signed comments will be strongly preferred (full name and valid e-mail address).

Leave a Reply

Your email address will not be published. Required fields are marked *

10 responses to “More on whether evolution really selects for veridical perception”

  1. aran arslan

    I do not think we can evaluate the contents of perception for truth or falsity since such contents are not propositions. I think on this matter Burge is clearly right. The veridicality conditions for perception can be set by accuracy or inaccuracy. The main role of the representational content is to set or to constitute the veridicality conditions of representational/psychological states. This constitution occurs in two ways depending on what kind of representational content it is (perceptual representational content or propositional one). The veridicality conditions can be set either by accuracy or truth. If it is a content of a perception, it constitutes veridicality through accuracy or inaccuracy conditions. If it is a content of a belief, it constitutes veridicality through truth or falsity conditions. The representational contents of perception cannot be evaluated for truth or falsity since they are not propositional.

    Reply
  2. Alex Morgan

    HSP's replies don't really engage with the substance of Cohen's points. In response to Cohen's point that HSP had originally mischaracterized the most plausible version of the veridicality claim, HSP emphasize that their primary target is not the view that individual perceptual states are identical to states of the world, but that *structural relations among perceptual states* are homomorphic to structural relations among states of the world. But if so, HSP simply aren't engaging with the view that most veridicality theorists actually endorse, namely the view that *individual perceptual states* accurately represent specific states of the world. They've just moved from one straw man to another. As Cohen emphasizes, to evaluate whether an individual perceptual state is accurate or veridical, we must specify what the content of the state is. HSP deny this on grounds that (i) information theory allows one to measure whether an information channel veridically passes messages without having to specify the content of those messages, and (ii) in their simulations natural selection pushes 'veridical' strategies in the direction of carrying little information about the world. But, again, this rests on a confusion about the relevant notion of veridicality. Information theory tells us nothing about whether an individual state is accurate; it tells us how *noisy* a channel is. Signals across noisy channels might nevertheless be accurate (consider a noisy phone line). More to the point, HSP's simulations can't tell us whether veridical strategies tend to carry little information about the world because they simply didn't model veridical strategies in the sense that veridicality theorists are concerned with. In that sense, a 'veridical' strategy would be one in which organisms tend to form perceptual states in response to the states of the world they're about, i.e. in response to their *content*. HSP didn't model the content of individual perceptual states, so their simulations provide no grounds whatsoever for the claim that veridical strategies carry little information.

    The other two points just trade on confusions stemming from the first two.

    Reply
  3. Jeremy Stafford-Deitsch

    Professor Hoffman argues that (1) evolution selects for fitness, not veridicality, and goes on to claim that (2) what we take to be actual objects in an external world are no more real than the icons of a computer screen. He then goes on to argue (3) that space-time itself is a similar illusory process we project onto an external world whose reality has no knowable connection to what we take to be real when we observe/ponder/investigate it. But if claims (2) and (3) are valid, how can he invoke entities/evolution/fitness – presuppositions that invalidate his conclusions – in order to reach conclusions that invalidate his presuppositions?

    Reply
  4. Jonathan Cohen

    glad to have an opportunity to say something about Hoffman's/HSP's replies, however briefly. taking their numbered points in order:

    1. moving from identity to homomorphism (though at odds w claims on p1484, but whatever) does avoid the implausible implication that perceptual states have to be identical to their distal targets in order to represent. nonetheless, the requirement that perceptual representation is a homomorphism between perceptual states and world states "that preserves all structures" (p1484) is still much stronger than proponents of representational accounts of perception have in mind. for example, the phenomenon of color metamerism, which HSP themselves mention on p1483, shows a gap between at least one kind of structure on the set of perceptual states and the set of the colors represented by the latter. (it's easy to generate other examples. e.g., the structure on visual states only preserves the size structure on world states up to a limit of perceptual resolution; presumably few proponents of representation take this to show that vision fails to represent distal size.) consequently, even if one took their arguments to refute the weaker homomorphism-required view, this would leave untouched the kinds of representational accounts of perception actually defended in the literature.

    2. we can put the logical point aside, since no one in this dispute thinks the perceptual contents under discussion are *formally* valid or contradictory. moving to information theory, it is true that one can use information theoretic tools without specifying the content of states. but to talk about information one needs to specify/individuate value ranges for the relevant independent variables. "one if by land, two if by sea" posits an informational relation between the number of lights (one or two) and the path taken by the enemy (by land or by sea). but that relation, and quantitative measures defined on it, crucially depend on the way we partition values: the quantity of information we say is in the signal depends on setting things up so that any sea route gets the same value of the variable for path taken. but that value partitioning question is exactly up for grabs in the setup of Hoffman's games (so not settled by them). is the variable measuring number of cans in a cell (so that 4 cans and 5 cans should get different values)? or payoffs for foraging in a cell (so that 4 cans and 5 cans should get the same values)? here, and in general, different such choices place different quantitative requirements on informational channels, and nothing about the game by itself or information theory forces a particular choice on us. it would seem, then, that the point (unsurprisingly) stands: one's assessment of the veridicality of the perceptual state depends both on what it says about the world (what does and doesn't count as an alternative) and how the world is.

    3. Hoffman says that his evolutionary games presuppose nothing, but rule out the sort of content assignments I proposed (e.g., the system represents payoffs in a cell as opposed to resources in a cell) on which they would turn out to be veridical. I agree completely that the games themselves fail to presuppose any particular content assignment. that is precisely why the dominance of one strategy over another in the game, taken by itself, doesn't tell you that the dominating strategy fails to involve representing the world veridically, or that the dominated strategy does involve representing the world veridically. (cf. point 2 above.) this is just to say that, Hoffman's claims to the contrary notwithstanding, the games do not rule out the sort of content assignments on which veridical representation is involved in the winning strategy.

    4. Hoffman insists that payoff are organism-dependent, and seems to infer from this fact that they can't be veridically represented by perception. that seems a non-sequitur. for instance, that I have a hand is an organism-dependent feature of the world, yet it is surely one that I can perceptually represent. if Hoffman thinks the very notion of veridical representation applies only to entirely organism-independent targets, then he is right to think that ordinary perception goes far beyond veridical representation. but, again, in saying this he's not disagreeing with proponents of a representational conception of perception.

    aran arslan: I think the Burgean point you articulate is orthogonal to the dispute between Hoffman and his opponents. as long as perceptual states have (and sometimes meet) veridicality conditions — whether constituted in terms of truth/falsity or accuracy/inaccuracy, then Hoffman is going to be unhappy.

    Reply
  5. Charlie Savelle

    On 2, he might say that a table is presented to us as an icon on a desktop might be, talking in terms of real may be begging the question a bit, I'm not sure. On 3, space-time isn't an illusion to him. To him it is a data structure. He defines an illusion as something without adaptive fitness.

    On the later point he claims that evolutionary theory is powerful in explanatory terms, and that is why he uses it, but that it is problematic because it posits things like genes. He says he uses evolutionary game theory, which only talks about abstract strategies and not space-time.

    https://youtu.be/oUWaEwbLkzQ

    Reply
  6. Tomas Orozco

    The low-hanging fruit of counterexamples to Hoffman's claims is curiosity. How can he rule out that the optimal division of cognitive labor in a population is achieved by true individual representations with intra- and inter- generational transmission of information? The optimality of curiosity would increase monotonically with the level of (both static and dynamic) uncertainty about fitness payoffs in the real world. If he can't rule out curiosity, he can't rule out veridical perception.

    Reply
  7. Jeremy Stafford-Deitsch

    For Hoffman's Interface Theory of Perception to work, one would surely need Hoffman's icons to have internal complexity and interrelationships as complex and intertwined as those biology claims to discover about the world (not much chance of that!).

    Hoffman's notion of spacetime 'predicts spacetime does not exist unperceived' (p.95 of THE CASE AGAINST REALITY). We must presume that our physical sense organs fall under the heading of icons which present pseudo-information about what we take to be the supposed middle-sized world. In chapter 3 of his book he discusses how cutting-edge physics supports his views – so suddenly we are back with veridicality, but for physics. But hold on. How can he argue that the sense organ icons fail with biological-level explanations (which he nevertheless wants somehow to keep with his evolution/fitness concept), but succeed at the level of physics? For consistency, surely one should argue either that they both work pretty well (in which case his non-Realist thesis fails) or that they both fail, in which case he can't use physics to bolster his position and we are in the untestable world of Berkeley.

    Reply
  8. Pendaran Roberts

    I read Hoffman’s popular press piece and much of his 2015 co-authored article in Psychon. Bull. Rev. If it has been determined (it hasn’t but more later) based on computer simulations that we are very unlikely to see the world veridically, then why should we think “veridically”? Do not similar arguments apply to our mathematical and logical faculties as apply to our perceptual faculties? Wouldn’t evolution select for thinking strategies that maximize payoffs not truths? And if this is correct, doesn’t the entire project collapse? We have somehow proven using math and logic that we cannot trust math and logic? Well, what a predicament we’ve gotten ourselves into! It is not enough to say as Hoffman does that “perhaps selection pressures favor accurate math” (p. 1500). It seems as if for the same reasons that we can assume it doesn’t, and whether it does or doesn’t can’t be tested. You can’t test whether a clock is accurate against itself or against another clock set by it. Further, it seems to me that whatever the rejoinder I’d be unlikely to be convinced. If, as Hoffman says, our perceptual language, involving space, time, objects, colors, tastes, and sounds, etc., is completely wrong to such an extent that there is no space or time nor objects, at least not as we understand them, then I see no reason to accept the arguments given by such creatures as Hoffman thinks we are about the nature of their reality. Such creatures, it would seem to me, should give up on saying what there is.

    Has it been determined that we are very unlikely to see the world veridically? The evolutionary games mentioned are very simple compared to the world as even a naïve subject understands it to be. For example, the robots (2015, p. 1987) could “see” two colors ‘red’ or ‘green’ and had to use these two colors to best represent their environment to maximize their fitness according to simple parameters. Unsurprisingly, the robots “evolved” to associate the colors with payoffs and not necessarily with the underlying structure of their reality (unsure why we even needed to run the simulation, as the result was obvious from the start). There is simply no reason for the robots to attempt to represent anything but the very simple payoffs. However, the real world is not like this. There are many different competing interests and needs of a species. The overarching evolutionary goal is to reproduce successfully, but that means keeping your offspring alive and aiding their development so they too can reproduce in a complex world. The more needs that a perceptual system must meet, the more likely it is to be tuned to the underlying structure of reality. For example, suppose the robots also had to save cans for the robot winter. In this system, the payoff structure would be more complex, and there would be a reason for the robots to differentiate 10 cans from 0 cans. Also, suppose the robots had a whole range of colors, not just 2. Obviously, we could create a system of payoffs according to which the robots represent reality in a way that is much closer to its actual underlying structure.

    A fortiori, whether a perceptual strategy is truth aligned depends on what the underlying reality is and on what the payoffs are. We can create simulations whereby the underlying structure and payoffs are such that the perceptions have hardly any alignment with the underlying truths. However, we can also create underlying realities and payoffs such that the perceptions align very well with the underlying reality and, in fact, even resemble this reality (to use Lockean language). Suppose that the robots instead of having 2 colors could “see” cans and numbers of cans. Suppose that the payoffs involved storing cans for later, that storing too many cans had a cost and that storing too few had a cost, and so on. It wouldn’t take long before we had created the conditions for a simulation whereby the robots perceptions aligned with reality.

    So, here’s the thing. The simulations show very little, in my opinion. You get out what you put in. The question is what should we put in such that the simulation correctly reflects the reality we are, in fact, in. And to this question there can be no non-question begging answer. What is the nature of reality? Well, that’s a philosophical question. We disagree for example on whether reality includes colors as we see them or just reflectances or on whether reality includes composite objects or just simples or on whether the reductive project succeeds or fails, and so on. Maybe only under a narrow range of realities and payoff structures would our perceptions be attuned to the underlying truth, but most of the realities are not plausible or even possible. There is a limited number of realities for which we have evidence and for which we can reasonably be said to occupy. If we do not want to assume that our perceptions are a reasonable guide to reality and so throw out all our evidence, throw out evolution too and also math and logic. In other words, a foundation of human inquiry is that we can gather knowledge about the world. If we are as far lost as Hoffman attests then there is no point in continuing the charade. Under the realities that we can reasonable be said to occupy, if evolution can be held to be true, then a priori it will be the result that our perceptions are reasonably accurate.

    Hoffman might try to say that we can continue with empirical inquiry but must understand it in idealist terms as really just self-inquiry: We are only really learning about the strange workings of our own minds. However, similar to what I’ve already said, if we are so far lost as this, then even this project is doomed. If our perceptual faculties are so far removed from reality as Hoffman claims then I see no reason to trust our abilities in any area. The idea that we could test the hypothesis that evolution would select for accurate math and logic won’t help. This can’t be tested, as math and logic are the foundation of any test. All inquiry presupposes that we can think accurately and if we can’t, then we can’t test whether we can. In other words, science and philosophy presuppose that we can have some faith in ourselves. Hoffman’s outlandish theory shows that we should have no faith in our own abilities, but then we should have no faith in Hoffman’s theory either. So, the entire project collapses in on itself. The foundations for rational thought do not allow some theories to be stated. Could they nevertheless be true? Probably not. The reality Hoffman is asking us to imagine isn’t conceivable. As Wittgenstein might say, Hoffman is breaking all the rules!

    There is another rather obvious issue with Hoffman’s position, which has already been touched on: Evolution, which is the foundation for his argument, is an empirical theory. If his view is correct, evolution is just part of the ideal world, not part of the real world. Whatever kind of world it is that we really occupy, a world of minds in a spaceless “void” according to Hoffman, there is no reason to think that evolution applies in this world, and according to Hoffman’s view, we have no access to this world perceptually. So, evolution cannot be an empirical theory for Hoffman but an assumption, and I’m not sure why we should make this assumption. Why should evolution apply to a spaceless world of minds; a world without genes, or reproduction, or even species. It seems he’s gotten himself into quite a mess! To end on a positive note though, I did find his papers very interesting. So, there’s something to be said for that! Good work Hoffman (and co-authors). I really enjoyed reading your work, but I don’t agree with it at all.

    Reply
  9. Don Hoffman

    I respond to Cohen’s first two points in order. I hope to have time soon to respond to the remaining two points.

    1. Cohen says that “the requirement that perceptual representation is a homomorphism between perceptual states and world states ‘that preserves all structures’ (p1484) is still much stronger than proponents of representational accounts of perception have in mind.”

    Our mathematical framework requires of world states only that they have a measurable structure (see Definition 2 of W). It requires nothing else—no specific other structures, or degrees of resolution, or freedom from ambiguity. Thus our approach and results apply broadly — not just to what we might take to be “our world” but to any “world” that can be discussed probabilistically. Such a world may have only one structure, or two structures, or a billion structures. It might have a small number of states and thus low resolution, or a great many states and thus high resolution. Our results still apply. Metamerism, limits of resolution, and so on are thus no problem. This generality of our results was a key reason we developed our mathematical framework, and is a key reason for mathematical models more generally.

    2. The domain of a payoff function includes states of the world. The states of the world may have structures — e.g., orders, groups, metrics, topologies. If a payoff function happens to preserve, i.e., be a homorphism of, such a world structure, then as natural selection tunes perceptions to the payoff function it will also happen to tune perceptions to be a homomorphism of the world structure—to be veridical. How likely is it that? One can show that the probability is zero (e.g., using counting measure for payoff functions that have finite domains and ranges, and then taking the limit as the domains and ranges go to infinity). Thus natural selection almost surely (i.e., except for a set of cases whose total measure is zero) does not tune perceptions to be homomorphisms of world structure, i.e., it almost surely does not tune perceptions to be veridical. Note that this theorem requires no assignment of content to world states or perceptual states. It is simply a fact about functions that depends only on the cardinality of the domains and ranges of the functions. No surprise. Theorems about functions don’t care whether you think about the domain and range as sets of apples or sets of oranges. The theorems apply in all cases. Such content is irrelevant to the theorem. Similarly, such content is irrelevant to our theorem that natural selection does not favor veridical (homomorphic) perceptions.

    Reply
  10. Jonathan Cohen

    I'm having trouble seeing how Hoffman's replies are responsive to the points made. again, taking his replies in order:

    1. I didn't deny the point Hoffman asserts — viz., that the mathematical framework is applicable to any sets of states on which a measure can be defined. what I did say was that going representational accounts of perception don't require the kind of structure preservation under discussion (which examples like metamerism and the resolution limits for visual size perception show by themselves), hence that using the general framework to argue that structure preservation is unlikely doesn't at all speak to the question of whether perceptual representation is likely or unlikely. the point is not that the framework is inapplicable, but that it doesn't show what it is claimed to show about veridical representation.

    2. grant that the simulations show (modulo worries about external validity etc) that homomorphism is unlikely. does that putative result show that veridical perceptual representation is unlikely? well, if veridical perceptual representation doesn't require homomorphism, then it does not. alas, no amount of insisting on the unlikelihood of homomorphism is going to help with this.

    —–
    KEYWORDS:
    Primary Blog

    Reply

Designed with WordPress