Todd and Gigerenzer begin their response by addressing the ideals that the adaptive toolbox conflicts with. These idealistic assumptions are:

1.      More information is always better – Sternberg argues “you are always better off using as much information as possible when time allows” for consequential decisions. Todd and Gigerenzer argue this ideal is misleading and “in order to make sound decisions in an uncertain world, one must ignore some of the available information”.  Engel points out that “frugality conflicts with legal systems, which often run on the defensive vision that more is always better” (767).
2.      Optimization is always better – Shanks & Lagnado argue that fast and frugal heuristics cannot account for human behavior because they do not optimize and human behavior can be optimal. Todd and Gigerenzer emphasize the distinction between optimizing processes (heuristics do not employ) and optimal outcome (heuristics can reach). Furthermore, using optimization does not guarantee an optimal outcome.  
3.      Complex environments demand complex reasoning strategies – Allen suggests that social environments are “responsive” rather than “passive” and therefore, are so complex that they require demonic reasoning abilities. Todd and Gigerenzer respond that “simple social exchange rules can coordinate social interactions without logical reasoning” (768). Furthermore, logic can make one predictable and therefore more open to exploitation.
4.      Search can be ignored – Oaksford argues “information is usually integrated in decision making by pointing to examples (such as speech perception and sentence processing) where the necessary information is available simultaneously, obviating search” (768).  Todd and Gigerenzer emphasize the requirement of searching for cues in many decisions situations.

Where do heuristics come from?

In this section, Todd and Gigerenzer address clarification questions about the origins of heuristics. They give three answers for where heuristics come from.

First, heuristics can arise through evolution. Baguley & Robertson express concerns about how heuristics get into the adaptive toolbox. Todd and Gigerenzer clarify they had no intention of giving the impression that evolution and learning are mutually exclusive. They further explain that “there would be strong selective advantages in coming into the world with a least some heuristics already wired into the nervous system” (768).  Houston mentions that there is growing literature surrounding the use of heuristic by animals in specific environments. Hammerstein states the difference between an optimizing process and an optimal outcome. He discusses evolution as an optimizing process for how decision mechanisms come about in biology. He also argues that “a heuristic that was evolved for adaptive use in one environment may be misapplied in another” (769).

Secondly, heuristics can arise through individual and social learning. Baguley & Robertson claim that certain heuristics can be learned. Todd and Gigerenzer agree with this claim but emphasize the distinction between adding a new tools to the adaptive toolbox and learning to use old tools in a new way. Solomon distinguishes two developmental changes that apply to this distinction. Changes in core “theories”, concepts, and reasoning principles may be called the tools in the toolbox while changes in “expertise” with age may change how or when tools are used. Furthermore, many developmental questions arise from the adaptive toolbox perspective. Barrett & Henzi question if fast and frugal heuristics are specific to human reasoning. Todd and Gigerenzer claim that humans rely heavily on social learning of heuristics which enables rapid addition of heuristics into the adaptive toolbox and is the foundation for culture. Gorman also describes how both heuristics and decisions themselves can be obtained from other individuals.  

Finally, heuristics can arise through the recombination of building blocks in the adaptive toolbox. Huber purposes that the adaptive toolbox is made up of many “partial” heuristics and multiple “partial” heuristics are used most decisions.Todd and Gigerenzer claim Huber overgeneralizes from the artificial lottery problem of risky choice and his assertion sets up a false opposition. The adaptive toolbox can hold both “partial” and “global” heuristics.  

How are heuristics selected from the adaptive toolbox?

Many commentators propose ideas about heuristic selection. Todd and Gigerenzer point out that “heuristic selection may not always be a problem” (770). Feeny claims that there are situations when heuristic selection is not needed. Heuristics in the adaptive toolbox are designed for specific tasks and this reduces the selection problem. Furthermore, the knowledge available reduces the set of possible heuristics when making a decision.

However, even after making the set of possible heuristics smaller, there still may be multiple heuristics applicable for a given situation. Morton suggests “a meta-heuristic which chooses between heuristics using the same principles as the fast and frugal heuristics themselves”. Feeney fears an infinite regress of meta- and meta-meta-strategies for determining the best heuristic and meta-heuristic for a given situation. Cooper and Feeney also worry that meta-heuristics won’t always pick the best heuristic. Todd and Gigerenzer respond with the claim that “the whole point of the adaptive toolbox approach is not aiming at optimization” (770). Furthermore, since more than one heuristic can be used in many situation, the choice between them isn’t critical.

Cooper questions the conditions required for selecting certain heuristics. Todd and Gigerenzer respond that most heuristic-relevant conditions haven’t been discovered yet. Margolis fears that heuristics that adapt to certain conditions may be harmful when significant changes to an environment occur. Todd and Gigerenzer point out that these maladaptive heuristics may be “replaced by new more appropriate heuristics acquired through individual or social innovation and learning” (772).

How can environment structure be characterized?

Todd and Gigerenzer stress the importance of environmental structure in order to understand cognition and behavior in terms of adaptation, as these are shaped by previous environments. Two approaches to studying the structure of environments are the search for environmental/ecological texture and the search for invariants in the visual environment. However, both these approaches overlooked the relationship between heuristics and environment. Simple Heuristics focuses on a few types of environment structures: environments in which lack of recognition is informative; noncompensatory information; scarce information; J-shaped distributions; and decreasing choice sets.

Allen and Sternberg touch on important points about cost-benefit analysis, stating that there are cases in which the cost of making a wrong decision outweighs the benefit of simple heuristics and therefore it is worth the extra time and analyses of the more consequential decisions. Todd and Gigerenzer respond by taking into account the significance structure of the decision, using the example of avoiding poisonous mushrooms differently when they are lethal or not. They conclude that while consequential decisions are often overthought and require justification, it is likely that the decision is reached very quickly, and much of the following thought processes are used to justify the decision that was made originally.

The difference between friendly and unfriendly environments, as Shanteau & Thomas point out, is that “friendly” envs. can all be positively correlated, while a subset of “unfriendly” envs. can be negatively correlated. Todd and Gigerenzer welcome this, further explaining that unfriendly envs. contain tradeoffs to simple heuristics. As the number of cues increases in an unfriendly environment, the performance of fast and frugal heuristics decreases. To mitigate this, we can combine partial heuristics to eliminate poor options first, then use a lexicographic process on the remaining.

An important area of adaptation to these heuristics is social rationality, and the authors agree with commenters calling for research to combine psychology and game theory for the understanding of interactive strategies. Simple heuristics may be exploited by more complex thought. They may also be exploited in J-shaped distributions, which occur in environments of power laws and Poisson processes.

Heuristics monitor and assess valid environmental cues at the surface level covarying with the decision variable rather than cues causally linked, since both are caused by the same process. It is possible a decision can be made with only one surface level cue, in comparison to the several or more casual cues necessary.

How can we study which heuristics people use?

It is difficult to obtain empirical evidence for heuristic use, but Todd and Gigerenzer stand firm in the belief that it is possible especially in the case of new predictions. In situations where there is little prior knowledge, it is easy to diagnose how they come to simple conclusions on new information. They criticize hypothesis testing in psychology, stating that there is no hypotheses other than a null hypothesis, and that we cannot assume how people use heuristics and use aggregated means. Instead, the adaptive toolbox is necessary to study heuristic usage, which includes (1) specifying potential heuristics, (2) deriving predictions for each, and (3) testing each participant’s judgement against those predictions. This encourages precise predictions based on individual differences in knowledge.

Solomon proposes that this methodology is important in ontogenic development, following pre-schoolers through childhood, adolescence, and into adulthood. Chater additionally speaks on the complementary studies of simple heuristics and rational analysis, which finds optimal solutions through elimination processes of simplifications.

What is the evidence for fast and frugal heuristics?

While it is widely accepted that animals use simple heuristics in their environments, it is more difficult to analyze human usage. Cooper and Harries & Dhami claim there is little evidence that humans use fast and frugal heuristics; however, Todd and Gigerenzer stand by their piece and empirical evidence along with other studies who found similar results. Humans often look up only one or two cues in order to avoid finding conflicting evidence. Cue dependency can be ignored in suitable environments for accurate strategies. Despite their hesitations, Harries & Dhami conclude that simple heuristics can even be used in legal or medical decisions.

The authors push back on Allen’s note that there is no evidence that people make inconsistent, intransitive judgments, using another study in which some people incorrectly inferred that A is greater than B, B is greater than C, BUT C is greater than A (which is intransitive).

Kainen asks for examples of heuristics in other fields such as perception, engineering, and language processing, but Todd and Gigerenzer stress that we must be aware of artificial mechanisms working similarly to the human mind.

What criteria should be used to evaluate the performance of heuristics?

Todd and Gigerenzer focus on multiple corresponding criteria rather than internal coherence. Since many people often violate first order logic (see the cities example) and ignore laws of probability, the authors question whether cognitive psychology properly evaluate human reasoning and cognition. Instead, they believe we should focus on correspondence between heuristics and environments.

Allen and Fuller note the tension between coherence and correspondence to the real world, despite the ability for a belief to satisfy both criteria. However, we cannot have complete coherence given our bounded rationality. Simple heuristics focus on inferences, where coherence and correspondence may diverge. Bermudez questions the justifiability of simple heuristics evaluated on success alone; the authors push back saying that the process of evolution and learning use success to improve processes for future decision making.  

Many responses extend the discussion of performance criteria of heuristics in societal and legal settings, raising fundamental questions. What are the implications of using simple heuristics with regard to the law?

Conclusion

This study of bounded rationality and its multidisciplinary results have relevance for many sciences attempting to understand organisms’ behaviors. Using the adaptive toolbox to find what lies inside is difficult, but fast and frugal heuristics allow us to reevaluate some underlying assumptions of world representations. We must extend our understanding of ecological rationality- how environment structures and heuristic mechanisms work together.

Questions:

Other than quick decisions, where and when can fast and frugal heuristics be effective?

It is likely that animals use fast and frugal heuristics most of the time. Humans obviously have capabilities much greater than this. However, are the fight or flight type decisions something that animals are superior to us in?

Where is the line between decisions and beliefs, especially if we are using simple heuristics

What are implications of using simple heuristics with regard to legal processes?

We will summarize the text,  offer our perspective on the author’s’ main points and pose some questions.

Unbounded Rationality

Todd and Gigerenzer propose two types of rationality: Bounded Rationality and Demonic (Divine) Reasoning. The latter assumes that “the human mind has essentially unlimited demonic or supernatural reasoning power” and is divided further into:

1. Unbounded rationality
2. Optimization under constraints.

Unbounded rationality represents a system that is omniscient and capable of certainty with unlimited access to information, however such a model is not realistic as the human mind has limited access to knowledge, time, money, and attention, when making decision.

The constrained optimization model requires a “stopping point” in the otherwise endless pursuit of knowledge. This point is reached when the costs of continuing to search for information outweigh the benefits of the additional knowledge that may be gained. This process also assumes almost unlimited time and computational power, as calculating the potential cost and benefit of each piece of further information would be hugely expensive and ultimately impossible.

The authors argue that “reasoning can be powerful and accurate without requiring unlimited time and knowledge”, which is why they move from the demonic model to the bounded model. Many philosophers and economists, however, argue that bounded rationality models “use a greater number of parameters and become more demanding mathematically” (730) than their unbounded alternatives, and for this reason they choose to stick with unbounded rationality. Todd and Gigerenzer respond to these critics by saying that “rationality need not be optimization, and bounds need not be constraints”(730).

Bounded Rationality

Thus, they present a notion of “Bounded Rationality” which presents the process of decision making as that of “a bounded [human] mind reaching into an adaptive toolbox filled with fast and frugal heuristics”(729). Bounded rationality accounts for both “the limitation of the human mind” and “the structure of the environment in which the mind operates”. The latter is a central theme of “Ecological rationality”, which shows the extent to which a heuristic is adapted to the structure of its operating environment. Subsequently, the authors present notions of “Satisficing” and “Fast and Frugal Heuristics”.

Satisficing is a method of making a choice from a set of options encountered sequentially when one “does not know much about the possibilities in advance”. It utilizes an “aspiration level” and stops searching for alternatives once an option that exceeds the aspiration level is found. However, difficulties arise during both the computation of the aspiration level and its relationship with each encountered option.

Fast and Frugal Heuristics limit the search of objects by using “easily computable stopping rules”, and make choices with “easily computable decision rules”. This represents Bounded Rationality in its purest form insofar as some fast and frugal heuristics make “satisficing sequential option decisions” and some do not. An example of a fast and frugal heuristic is One Reason decision making where one single piece of information is used to make a choice. This sharply contrasts with heuristics of unbounded rationality which attempt to assimilate unlimited information in different decision making environments.

The authors explain the fast and frugal heuristic principles that:

1. Guide search
2. Stop search
3. Make decisions based on (1) and (2)

Regarding Search guidance, fast and frugal heuristics use ‘simple search strategies’, such as cue orderings, which are easy to compute and ignore dependencies between cues.  This contrasts with ‘active search strategies’ which use extensive computations or knowledge to find the next cue or alternative to search.

Regarding Stopping search, fast and frugal heuristics employ simple rules. For example: cease searching and make a decision as soon as a reason that favors one alternative is found. For alternatives, simple aspiration level stopping rules can be used.

Regarding principles for decision making(3), decisions can be based on only one reason, regardless of the total number of cues found during search. Also, cues are not weighed or combined into a “common currency”. Rather, a simple elimination process is employed until one final choice remains.

The authors note that the components of fast and frugal heuristics are constructed by combining building blocks, such as One Decision making, or nesting existing heuristics within each other, such as a Recognition Heuristic that works on the basis of an elementary cognitive capacity and recognition memory.

Fast and frugal heuristics are divided into four classes, which are :

1. Ignorance Based Decision making
2. One Reason Decision making
3. Elimination Heuristics for Multiple-option choices
4. Satisficing Heuristics

First, Ignorance based decision making ensures that, in two-option situations, decision makers choose recognized options over unrecognized options. The authors give an example in which humans are more likely to choose recognizable ham over “odd colored” eggs. Such ignorance based recognizing heuristics perform better than random choice in environments where a positive correlation exists between the decision maker’s exposure to different possibilities and the ranking of these possibilities along the decision criteria being used. In other words, the more recognizable an option is, the more likely it is that it will get chosen.

Second, One Reason decision making ensures that, when multiple cues exist, a stopping rule is activated once enough information has been gathered to make a decision. In other words, the decision maker stops looking for cues as soon as a cue is found that differentiates between the options considered. Hence, such heuristics include both a stopping rule, and a decision cue. Moreover, cues can be searched by the following methods:

1. i) In order of validity (Take the Best heuristic)
2. ii) According to the order determined by their past successes in stopping search

iii) In random order (Minimalist heuristic)

The authors note that such heuristics are non-compensatory as “once they have used a simple cue to make a decision, no further cues in any combination can undo or compensate for that one cue’s effect”.

Third, Elimination Heuristics for multiple option choices are employed when each available cue “dimension” has fewer values than the number of available alternatives. One Reason decision making is insufficient in such cases because a single cue cannot distinguish between all alternatives. For example, knowing whether or not 15 cities contain a river is not enough to decide which is the most habitable. Moreover, elimination heuristics employ both stopping rules and elimination. The former ensures that cues are only sought until enough is known to make a decision while the latter ensures that successive cues are used to eliminate options until a single option remains.

Fourth, unlike other heuristics that assume all possible options are available to the decision maker, Satisficing heuristics find alternatives sequentially over an extended period or spatial region. This is done by setting an aspiration level and the search for alternatives is stopped once the selection criterion is met.

The authors posit that organisms deal with their environments. Hence, different environments can have different specific fast and frugal heuristics that exploit the information structures present to make adaptive decisions. However, this could lead to a multitude of specific fast and frugal heuristics. Thus, the authors propose the following information structures which make fast and frugal heuristics “ecologically rational”:

1. Non-compensatory Information
2. Scarce Information
3. J-shaped distributions
4. Decreasing Populations

Regarding Non-compensatory information, the potential contribution of each new cue falls rapidly. In such cases, the Take the Best heuristic is ecologically rational.

Regarding Scarce information, the Take the Best heuristic outperforms  a class of linear models on average when few cues are known relative to the number of objects.

In J-Shaped distributions there are many small values and few large values. In such cases, the QuickEst heuristic, which takes values of objects along the same criterion using as little information as possible, is ecologically rational.

Regarding decreasing populations where the set of options is constantly shrinking, a satisficing heuristic is ecologically rational.

Furthermore, the authors propose that fast and frugal heuristics should be evaluated based on correspondence criteria regarding speed, frugality and accuracy. This contrasts with coherence criteria which are primarily concerned with the internal logical coherence of judgements. Moreover, this allows for the comparison of heuristics within the actual requirements of its environment which may include, using a minimal amount of information or making accurate decisions in a minimal amount of time.

Lastly, the the authors conceive of the mind as an “adaptive toolbox” which is a collection of specialized cognitive mechanisms that evolution has built into the human mind for specific domains of inference and reasoning. These contain psychological adaptations and include perceptual and memory processes. However, higher order cognitive processes are better modeled by simpler algorithms than lower order mechanisms.

Conclusion

The authors conclude by stating the following challenges to be addressed by their research:-

1. Cognitive tasks such as planning, perceptual mechanisms, and problem solving
2. Adaptive problems regarding domain specificity and the organization of the mind
3. Social norms and emotions
4. Ecological irrationality regarding the aspects that shape the design and performance of decision heuristics
5. Performance criteria, that is, is there a role for coherence criteria?
6. Selecting heuristics, in other words, how does the mind know which heuristic to use?
7. Multiple methodologies regarding the prevalence of simple heuristics in human and animal reasoning

The authors’ conception of fast and frugal heuristics, although probable, seems to be lacking in both empirical evidence and in the complexity of the human decision making process. Regarding the former, it is acknowledged that empirical means of testing search processes during human decision making are either non-existent or not utilisable. This is problematic as it leaves one wondering what theory of knowledge and categorization of epistemology fast and frugal heuristics adhere to. Regarding the latter, they seem to overlook the complexity of nesting heuristics, calculating the validity of cues, and the distinction between conscious and unconscious rationality, if existent. Lastly, one is forced to ask whether the authors are trying to understand cognition in the context of modeling rational systems, and/or understand how humans reason. In other words, do humans inherently embody ecological rationality in ways that artificial systems cannot (example: early vision and visual recognition).

Questions

1. Does this paper serve as an argument in support of the naturalisation of epistemology? If so, is it sufficiently supported by empirical findings? If not, is it supported by an externalist, coherence, or foundationalist theory of knowledge?
2. At what point does the nesting of different fast and frugal heuristics require extensive computation, time, and knowledge? Think about making simple vs complex decisions
3. How do we measure the validity of a cue in relation to other cues? Do Todd and Gigerenzer oversimplify the cue selection process?

1. The General Approach

1.1 Philosophy, Psychology, and Artificial Intelligence

Paul Thagard’s response to the critiques of his article “Explanatory Coherence” begins with a discussion of the criticisms that Dietrich and Wetherick pose about categorizing Thagard’s theory of explanatory coherence (TEC).  Dietrich mentions his confusion in distinguishing between whether TEC is a theory of philosophy of science or of psychology. Thagard explains that ideally TEC belongs in both, but this may not always work out perfectly.  Dietrich concludes that it makes slightly more sense for TEC to be a psychological theory because he does not see a point to ECHO if TEC is a theory of philosophy of science.  Thagard clarifies that the point of ECHO is to provide a more in-depth example of TEC than has previously been supplied.  In addition, Thagard notes “TEC and ECHO…are not positivist because the emphasis in on high-level theories, not on observation, and data can be rejected; and the principles of explanatory coherence go well beyond formal logic” (491).  Thagard coins the term “biscriptive”, a combination of descriptive and prescriptive, which describes how people make decisions when in accord with the best options available, therefore offering insight as well as a critique into how human performance functions.

Wetherick does not think that one theory can be applicable to both psychological processes as well as to sociological ones, and that TEC is sociological.  Thagard recalls his explicit use of Lavoisier and Darwin as models for ECHO, and that his “examples come from the history of science, not its sociology.  Nor do I pretend to model minds free of prejudice and preconceptions” (491).  This is also a response to Wetherick questioning the success of a connectionist model when emulating a prejudiced mind.  Wetherick argues that conscious symbolic processing always determines perceived explanatory coherence, and that Thagard does not include symbolic processing at all in his model.  Thagard responds by mentioning that people often find coherence after they have stopped consciously thinking about it, and therefore explanatory coherence is not always determined by symbolic processing.

1.2 Connectionism

Thagard acknowledges that the ECHO relies mainly on connectionism to demonstrate TEC, but Dietrich believes that ECHO is not directly related to TEC.  Thagard argues that ECHO is vital to TEC because it offers examples to back up this theory and that TEC evolved from the creation of ECHO.  Lycan questions Thagard’s heavy use of connectionism by arguing that connectionist architecture has no advantage over von Neumann architecture and that ECHO’s use of connectionism is actually not a strong trait in the model because it could achieve the same results by using a traditional architecture.  Thagard responds that connectionist models act as supplements to traditional architecture.  A nonconnectionist version of ECHO can be used, but the connectionist model results in more appropriate conclusions.

2. Theoretical Issues

2.1 Explanation and hypothesis evaluation

How do we define explanation?  Achinstein wonders if we can use Thagard’s theory successfully without also having an operating definition of explanation, to which Thagard expresses his current incapacity in creating a theory that incorporates all types of explanations.  Achinstein also describes how there needs to be a connection between explanation and acceptability.  Thagard retaliates how “the inference from ‘H is the best explanation of the evidence’ to ‘H is acceptable’ does not require any special relation between explanation and acceptability” (492).

O’Rorke brings up the interesting issue of how an agent’s goals and priorities are important factors in evaluation, yet Thagard thinks that these factors only matter for the generation of hypotheses and not whether an agent will accept a hypothesis or not.  Thagard mentions his desire to create MOTIV-ECHO, a program that will “reach conclusions on the basis of how well beliefs satisfy its goals, as well as on the basis of how much explanatory coherence they have” (493).  Thagard also describes how people tend to supplement their own desired conclusions with evidence that they have selectively applied to a situation, and will ignore evidence for an unfavorable conclusion.  On a similar note, Sintonen recognizes that people often make choices because of the promise they show, not necessarily because they have coherence.  Thagard agrees and adds that because of this, people do not always choose the best available option.  In developing MOTIV-ECHO, Thagard hopes to incorporate this phenomenon as a part of the program’s ability to rationalize.

2.2 Simplicity

Reggia emphasizes and exploits the claims of simplicity, which Thagard defends, to suggest the alternative use of Bayesian methodology. Thagard confidently responds that this interpreted ‘principle of cardinality,’ where “explanatory hypotheses with the smallest number of hypothesized components are preferred” (p. 493), is not always the case and instead has more so to do with the configuration of hypotheses. Furthermore, he states that frequency and probability-based hypotheses are not always found in the scientific and legal domains to which ECHO is applied, and therefore may not function as the best alternative for TEC.

2.3 Analogy

            While Thagard believes that analogies may improve the coherence and acceptability of hypotheses and theories, McCauley disagrees by saying that theories have a role in determining the nature of analogies and thus cannot be evaluated in this way. Similarly, Hobbs argues that these analogies improve the explanatory power of theories that rest on abstract principles. Such theories and abstract principles, however, do not have any influence over analogy recognition, which according to Thagard & Holyoak, can be achieved through the examination of pragmatic, semantic, and structural constraints. Although ECHO only addresses pragmatic constraints, many other theories (i.e. natural selection) appear to address individual constraints depending on their intended use of analogies. As for analogical reasoning, Gabrys & Lesgold point out the distinction between the case-based reasoning of juries and judges. Despite their intentions, this critique offers no objection to the constraint-satisfaction model, which we find in law and computational programs, like ECHO and the analogy program, ACME.

2.4 Conceptual change

            Many critics, like Giere, attack Thagard and TEC for its lacking intentions to adjust and account for historical change in science. Giere questions whether Thagard’s model truly represents the reasoning of scientists and whether it can explain the transition from old to new theories. Thagard does not provide a clear answer to these limitations, but hopes to further examine the breadth of ‘judgmental’ and ‘representational’ mechanisms used by practicing scientists to acquire understanding of explanatory relations over time until new theories form. Nevertheless, Thagard believes that psychological experiments are of utmost importance in understanding and representing human reasoning.

Mangan & Palmer, in contrast, adhere to the Kuhnian idea that scientific revolution changes the methodological principles of explanatory coherence faster than TEC can account for them. Thagard argues that these philosophers grossly exaggerated the historical variability of such principles. To support their argument, Mangan & Palmer describe Darwinian revolution as a shift in the methodological use of analogy to improve theoretical explanatations. Thagard, however, finds evidence of this shift in the writing of philosophers and scientists (Paley, Huvgens, & Fresnel) well before Darwin’s time thus contradicting the Kuhnian dogma.

2.5 Logic & Probability

Other critics, like Feldman, begin to challenge the semantic foundation of TEC in comparison to probability theory. Although examples of “clean and well-understood formal semantics” may be hard to find and incorporate into ECHO, logic and probability theory should have no greater success (p. 495). Thagard points out that the Tarskian semantics of predicate calculus avoid the ‘central semantic question’ by repeatedly assigning definitions to progressively complex formulas. Furthermore, probability is neither fully understood nor interpreted in terms of its definition and theoretical axioms.

Cohen emphasizes the importance of predictions, which when successful, may signal simple explanations to hypotheses without post-hoc additions to the original theory. He also criticizes TEC and ECHO for having no way to “determine the acceptability of a conjunction based on the acceptability of the conjuncts” (p. 495). While Thagard admits this to be true, he reminds us that probability theory fails to a similar degree in calculating conjunctive probability from the indeterminate level of dependence between conjuncts. Such probabilistic knowledge is rarely available to the domains of ECHO. Despite these quantitative limitations, Dawes uncovers ECHO’s ability to deal with the conjunctive complexity of Simpson’s paradox where supporting evidence may refute a hypothesis if both taken under consideration.

3. Problems with the ECHO Model

Many commentators expressed problems with the ECHO model and suggestions for improvement. A recurring issue is the arbitrariness of inputs to ECHO when considering explanations versus analogies or hypotheses versus evidence. McCauley questions how disputants can neutrally decide what constitutes an analogy because virtually anything can be analogous. Thagard claims this skepticism is only relevant if the Kuhnian view, which implies that fundamental principles are malleable, is correct. However, to him, this view is exaggerated. Additionally, both Dietrich and Zytkow question ECHO’s ability to distinguish between evidence and hypotheses, but Thagard doesn’t find the distinction problematic either.

Bareiter & Scardamalia notice the problem that when competing hypotheses are inputted into ECHO, they can affect the weight and activation levels of each other, and therefore result in differing outputs. Thagard explains two ways to address this: enrich the input to notice contradictions that have already been omitted, or adjust the output threshold because once a unit hits subthreshold, it no longer affects the activation of others. Meanwhile, McDermott questions ECHO’s output in general, stating that it may be established from the weight of links between propositions based purely on the problem’s structure instead of its content. Thagard dismisses this as peripheral and states that the model as a whole shows how explanatory hypotheses can be examined in complex ways.

Finally, commentators present alternate methods to compare to ECHO. Hobbs describes a “naïve method” of counting propositions which subtracts the number of hypotheses a theory uses from the number of evidence pieces it explains (#E – #H). This method, Hobbs claims, arrives at the same conclusion as ECHO, so why bother using ECHO at all? Thagard then claims that there are numerous possibilities in which ECHO and the naïve method don’t yield the same conclusion, and therefore are not equivalent. Simon and Zytkow introduce the connectionist model STAHL and compare its abilities to ECHO. Thagard acknowledges STAHL’s abilities, but doesn’t agree that it is a more comprehensive model of theory evaluation. STAHL is a discovery program and therefore more effective than ECHO in considering proposition content, while ECHO models evaluation, so is superior to STAHL in that it’s not restricted to a single method of hypothesis evaluation. Thagard eventually acknowledges that collaboration of models could produce an improved method of theory evaluation.

4. Psychological Adequacy

Many commentators question the psychological adequacy of TEC and ECHO. First, as Klayman & Hogarth state, the examples utilized in the target article to exhibit the use of ECHO in TEC are insufficient tests of psychological validity. They state that ECHO doesn’t model the process of thinking, but rather its end result, and therefore is incapable of representing cognition (478). Thagard asserts that researcher have begun launching experiments to grasp the empirical side of ECHO and support its psychological adequacy. Yet many authors, specifically Earle, express concerns about the testability of ECHO due to the subjective nature of what is considered rational when credible hypotheses conflict.

Cheng & Keane state Thagard’s account is too holistic and parallel to be deemed psychologically adequate. Individuals approach theory evaluation in a more piecemeal fashion than ECHO, which considers all hypotheses and evidence of a theory simultaneously. Thagard asserts that although this may not be within human capacity due to limitations in short term memory, most evaluations of the explanatory coherence of propositions happen unconsciously, and therefore represent parallel and holistic judgment. Therefore, although a cognitive model would integrate ECHO with conscious deliberation, Thagard believes this integration can occur.

Finally, several commentators (Chi, Bereiter & Scardamalia, and Read & Miller) suggest the use of ECHO to explore certain psychological phenomena, including conceptual changes within individuals, whether people can learn to evaluate hypotheses more effectively (and to encode pieces of evidence they disagree with), and the application of TEC to social phenomena. Thagard is enthusiastic about these approaches, yet notes that more research is needed in psychology, philosophy, and AI in order to push forward.

Discussion Questions:

1. Is a working definition of “explanation” needed in order to efficiently evaluate TEC?
2. Should we use a program like ECHO to ‘teach’ people to evaluate hypotheses more effectively in a rational way?  What would be the implications of this in society, science, and law?

Introduction

Paul Thagard’s article “Explanatory Coherence” discusses a negative coherence theory based on seven principles. These principles establish relations of local coherence between hypotheses and other propositions. Thagard presents these principles through their implementation in a connectionist program called Echo. In Echo, coherent and incoherent relations are encoded by excitatory and inhibitory links, respectively. Echo’s algorithm is based on considerations of explanatory breadth, simplicity, and analogy. Thagard argues that Echo simulates human reasoning by  accepting or rejecting hypotheses.

Explanatory Coherence

Thagard defines explanatory coherence as propositions “holding together because of explanatory relations.” (436) The author’s definition of explanatory coherence is based on: “(a) a relation between two propositions, (b) a property of a whole set of related propositions, or (c) a property of a single proposition” (436). Thagard claims that “(a) is fundamental, with (b) depending on (a), and (c) depending on (b)” (436). The author focuses on the acceptability of a proposition, which is independent of a set of propositions. The greater the coherence of a proposition with other propositions, the greater its acceptability. Thagard also states that although explanation is sufficient for coherence, it is not necessary. For instance, two prepositions can cohere for nonexplantory reasons, such as in deductive, probabilistic, and semantic coherence. Incoherence occurs when two propositions contradict each other.

Thagard establishes seven principles as the makeup of explanatory coherence:

1. Symmetry: asserts that pairwise coherence and incoherence are symmetric relations.
2. Explanation: states (a) “what explains coheres with what is explained” (437);  (b) “that two propositions cohere if together they provide and explanation” (437); and that (c) theories with fewer propositions are preferred.
3. Analogy: states that (a) analogous propositions cohere if they are also explanatory; and (b) “when similar phenomena are explained by dissimilar hypotheses, the hypotheses incohere” (437).
4. Data Priority: assumes that each proposition contains a degree of acceptability because they were achieved by methods that led to true beliefs (464); if the preposition  doesn’t fit with other beliefs, then it has to be “(un)explained away.” (Professor Khalifa)
5. Contradiction: syntactic and semantic contradictions.
6. Acceptability: proposes that (a) “we can make sense of the overall coherence of a proposition in an explanatory system from the relations established by Principles 1-5” (438); and that (b) when a hypothesis only accounts some evidence in the system, it is a weak hypothesis.
7. System Coherence: local coherence of propositions dictate the explanatory coherence of the whole system.

A Connectionist Model: Echo

Thagard introduces connectionist models to aid the reader’s understanding of Echo. Connectionist techniques describe networks in terms of units that excite or inhibit other units. For example, when looking at the Necker cube, such that A is a corner on the cube face in the foreground, then one must also perceive corners B, C and D as on this face (439). This focusing of attention on A is termed “activating” A. The relationship between A and the other three corners demonstrates that A excites B, C and D. Ultimately, the connectionist model reveals a holistic approach to perception (438).

Thagard advocates that Echo simulates human reasoning. Consistent with other connectionist models, Echo is based on excitatory and inhibitory links: if Principles 1-5 state that two propositions cohere, an excitatory link is established between them; if they incohere, an inhibitory link is established. Echo attributes the same weight to propositional links because they are symmetric (Principle 1). However, if an explanation has a larger number of propositions, the degree of coherence between each pair of propositions decreases, and Echo proportionally lowers the weight of excitatory links (Principle 2). Hypotheses that explain analogous evidence cohere with each other (Principle 3).  When the network is run, activation spreads from a special unit that always has an activation of 1, giving each unit acceptability (Principle 4). Units that have inhibitory links between them have to compete with each other for activation. The activation of one of these units will suppress the activation of the other (Principle 5) (439).

Echo’s values are set by a programmer, therefore it is not objective.  Specifically, the variability of Echo’s parameters (tolerance, simplicity impact, analogy impact, skepticism, and data excitation) appear arbitrary. However, Thagard insists that if a fixed set of default parameters apply to a large range of cases, then the arbitrariness is diminished (443).

Applications and Implications of Echo

The reading describes several examples of how Echo supports scientific theories and legal reasonings. However, the Craig Peyer trial is particularly interesting because it is the only provided example where the findings of Echo were inconsistent with the actual outcome. The jury’s decision was not unanimous in Peyer’s trial. Echo, however, found that it was easier to reject Peyer’s innocence hypothesis than that of Chambers (452). Ultimately, this example demonstrates a disconnect between Echo and actual human reasoning.

Thagard explores how connectionist models compare to other programs used to develop artificial intelligence (AI). Echo’s major limitation on AI development is that it only represents one method of exploring rationality and cognition (457). Thagard encourages collaboration across many disciplines, such as neuroscience and mathematics, in order to fully understand the mind. He specifically contrasts connectionist models with probabilistic models and explanation-based learning. Thagard acknowledges that probabilistic models are attractive because they are based on axioms. However probabilistic models cannot evaluate scientific phenomenon that do not have a statistical foundation (459). Furthermore, connectionist models may resolve issues with explanation-based learning in AI. Some explanation-based learning systems perform hypothetical experiments to identify causal relationships, but this method is not practical for complex theories. Connectionist models can enhance these systems by offering comparisons between multiple explanations to select the strongest relationship (459).

Thagard also discusses the role of connectionist models in relation to psychology, specifically in attribution theory, discourse processing and conceptual change (459-461). He highlights that Echo models how individuals attribute information about their surroundings to develop causal explanations for their observations (460). However, Echo cannot simulate how individuals consider alternative reasons for another individual’s behavior, such as that they were coerced (460). Thagard commends Echo for mapping human discourse processing, including interpreting a question as a request or decoding a story. He argues that individuals are constantly evaluating hypotheses about others’ intentions and meanings in texts (460). Lastly, Thagard praises Echo for simulating the shifts in beliefs of subjects learning a new phenomenon when they are given more evidence to aid their understanding (460).

In terms of their implications in philosophy, connectionist models contradict with several other theories. Echo employs a holistic approach because it evaluates a hypothesis in relation to other propositions within a larger system. However, Echo contradicts with holism in that it can also consider local propositions (463). Second, Thagard acknowledges that Echo contradicts with probabilistic theories. Echo does not define propositions by their support in probability axioms, but by degrees of activation. Similar to his discussion of Echo on probabilistic models in AI, Thagard claims that some concepts cannot be assigned a probability (464). Third, Thagard asserts that explanatory coherence contradicts with confirmation theories because these theories evaluate a hypothesis based on the number of observed instances of its claim. He believes this method has similar limitations to the probabilistic models used in psychology and philosophy (464). By considering Echo’s implications on various fields of study and real world scenarios, we may better understand its advantages and weaknesses.

Critiques of Echo

Thagard acknowledges that Echo’s limitations consist of a programmer bias, a lack of a definition for “explanation,” the normative and descriptive dichotomy, and dependence on causal relationships and the seven principles. As previously discussed, a significant limitation of Echo is that the programmer encodes the initial data. Therefore, this individual has agency in how the claims are framed, which might reduce the reliability of the outputs (453). Another restriction is that Echo is based on explanatory relationships between propositions, but there is not a consensus on the definition of an “explanation.” Furthermore, Thagard advocates for viewing explanatory coherence as both descriptive and normative, or describing how people reason and how people should reason (465). It is unclear whether Echo is descriptive or normative. Additionally, Echo can only analyze causal relationships (454). Lastly, the quality of Echo is dependent on the quality of the seven principles. These principles have the potential to be disproven or expanded upon, therefore Echo in its current state might not be the best version of itself (456).

Questions

1. What are the advantages of using Echo’s and its parameters?
2. Thagard supports explanatory coherence on the basis of the theory’s connection with acceptability. How can we be sure that an explanation is acceptable? What is an explanation?
3. In supporting a hypothesis, Echo distinguishes between the strength of propositions and the amount of propositions? How do humans evaluate this difference?
4. Consider the example of Craig Peyer’s trial and how the outcome of Echo was inconsistent with the outcome of the jury. What differences does this example suggest between Echo and human rationality? Is Echo a prescriptive or descriptive model?

A potential paper topic is the following:

How (if at all) does psychological research concerning the cognitive impenetrability of vision increase or decrease the plausibility of foundationalism as a theory of justification?

In what follows, I sketch some of the most plausible and interesting arguments that I can think of that are relevant to answering this question. My goal here is not to take a stand, but to present some potential ways of thinking through these issues. It’s up to you whether you want to develop these arguments, or come up with your own arguments. (Remember that you can also write about Almeder vs. Churchland or about coherentism, which we’ll discuss next week.) You should feel free to challenge the premises of any argument offered here, and also think about how the arguments relate to each other.

If you plan on using any of these arguments below, make sure that you can motivate each of the premises. Even if you ultimately reject the premise, you should first consider why would a reasonable person find it plausible in the first place. This will help you when you have to walk through these arguments.

First Argument: penetrability falsifies foundationalism

Arguably, here is the clearest way to use the psychological literature with respect to foundationalism:

1. Vision is cognitively penetrable.
2. If vision is cognitive penetrable, then visual beliefs are justified only in relation to other beliefs.
3. If visual beliefs are justified only in relation to other beliefs, then foundationalism is false.
4. :. Foundationalism is false. (From 1-4)

Obviously, Pylyshyn has given us many reasons to question Premise 1. So, if you want to advance this argument, make sure that you can rebut his toughest objection to this premise.

There might also be grounds for challenging Premise 2. What exactly is a visual belief? Also, maybe the psychological literature favoring penetrability only shows that visual beliefs require other beliefs for something other than justification. If so, you should state clearly what this “non-justificatory aspect of vision” is, and explain why it is not justificatory.

Second Argument: impenetrability supports foundationalism

Note that even if this argument is unsound, it doesn’t follow that cognitive impenetrability favors foundationalism. In particular, we might see Pylyshyn as only arguing this much:

1. If vision is cognitive penetrable, then early vision can access relevant expectations, knowledge, and utilities.
2. Early vision cannot access relevant expectations, etc.
3. :. (Early) vision is cognitively impenetrable. (1,2)

We can then see much of his target article as providing detailed arguments for Premise 2, and some of the commentary as challenging this premise. However, those commentators who challenge the very distinction between early vision and cognition might also be challenging Premise 1.

But even if we can work out this argument, how do we get from cognitive impenetrability to foundationalist theories of justification? This requires a further argument:

4. The (non-motor) outputs of early vision are mostly veridical.
5. If the (non-motor) outputs of early vision are mostly veridical and early vision is cognitively impenetrable, then early vision provides a foundation for our knowledge.
6. :. Early vision provides a foundation for our knowledge (3,4,5)

Pylyshyn argues for Premise 4 (largely when discussing “natural constraints.”) Premise 5 seems plausible to me, and, as I mentioned in class, and “mostly veridical” is a very attractive criterion for a basic belief. It’s not as restrictive as incorrigibility, but it’s not as flimsy as prima facie justification.

However, Premise 5 will require some further clarification. In particular, what exactly are the non-motor outputs of early vision? Look this up in Pylyshyn (especially in the section of his response called, “Where is the boundary of early vision?” and the commentaries to which that section refers). Also, what’s “a foundation of knowledge”? Given our goal of putting the psychology into the service of epistemology, there are two big questions we need to answer:

1. Are some (non-motor) outputs of early vision beliefs? In other words, are the non-motor outputs of early vision the kind of things that can be true or false? Or can they only be more or less accurate? (Compare: the belief that a square is in front of a circle can be true or false; however maps are neither true nor false, they are more or less accurate.) If they’re beliefs, then you have basic beliefs! If not, you should challenge the doxastic assumption. Note that Pollock and Cruz already give you a story about non-doxastic versions of foundationalism—parts of it are in Chapter 1, and other parts are in Chapter 2.
2. How much/what kind of information do these outputs convey? Pylyshyn seems to suggest that they are restricted to geometric information. (See if he says more.) But this raises a problem akin to those raised in the “Epistemic Ascent” section of the Pollock & Cruz chapter on foundationalism: will we be able to justify beliefs about physical objects on the basis of a purely geometric output? How can we instantiate arguments (in the Pollock and Cruz sense) from this geometric output all the way up to physical object beliefs?

Third Argument: penetrability is irrelevant to foundationalism

Finally, perhaps inspired by Almeder, you might think that psychology and epistemology are mostly independent of one another. Arguing for this is a bit trickier. Here are two strategies:

First Strategy: Go through the two preceding arguments and find faults with both of them.

Second Strategy: Show that all four of the following are possible:

1. foundationalism+impenetrability,
2. foundationalism+penetrability,
3. coherentism+impenetrability,
4. coherentism+penetrability.

Obviously, this will be easier to do once we study coherentism.

Introduction

Pylyshyn’s article “Is vision continuous with cognition? The case for cognitive impenetrability of visual perception” (1999) received a large number of commentary responses, and although some commentators spoke in agreement with Pylyshyn’s claim that one aspect of vision (“early vision” as dubbed by Pylyshyn) is independent of cognitive penetration. He observed that, of the 40 some odd commentaries, there seemed to be three general response themes. 1) There really is no point in distinguishing vision from cognition, either wholly or partly, for a host of reasons ranging from unimportance (with the argument that consciousness or conceptualism are more important arguments) to the notion that it will always be impossible to decide or prove whether or not Pylyshyn is correct or incorrect; 2) Our current understanding of neurophysiology does not fully support Pylyshyn; 3) Pylyshyn’s idea is mostly valid, except that its points of reasoning may not be as tight as they could be if drawn from other points. Pylyshyn continues to, conveniently, address his commentators in the following six sections.

Distinctions and decidability

Pylyshyn defends his premise that vision is a simplistic structure: it’s binary, and an input-output system (402, 404). While Uttal and Dresp argue that vision cannot possibly be as simple as input and output due to the complexity of the brain—which also makes it immeasurable—Pylyshyn attests that even complex structures can be broken down and that, at this time, we may not have the tools to measure the visual pathway’s binary-ness (402). But that does not mean the empirical science will not one day be developed and the premise will be decidable. As for vision as a binary, Pylyshyn asserts that a modal system would require the representations and knowledge developed by the visual system to make a modal model, yielding cognitive penetration (404). That visual “representations require representation-interpretations processes” (403) is a circular argument that Pylyshyn disassembles with his idea that representations do not occur yet in early vision because representations are cognitively penetrable. Likewise, he asserts that visual apprehension is not knowledge-dependent (as Pani and Schyns suggest) because of the neural pathway it takes; early-vision itself does not interact with cognitive pathways. Rhodes and Kalish’s method of distinguishing early vision from post-visual decision, Signal Detection Theory (SDT), is not effective enough to support or deny Pylyshyn’s argument (404). Additionally, the color testing proposed by Schirillo does not control for the cognitively penetrable effects of memory, and therefore nullify his claim that color memory penetrates early vision (404-405).

Top-down effects versus cognitive penetrability

Once again, Pylyshyn fields heavy criticism that he underestimates the complexity of the brain. He replies to Tsotsos, Grunewald, and Bullier’s critiques that early vision is not exempt for top-down effects by asserting that in a complex system, of course there are “interlevel influences” (405), but these cognitive biases are formed before or after early vision and therefore are not pure early vision and cannot affect the raw perception of stimuli. Grunewald continues to argue that delayed stimuli cognitively penetrate early vision (406), although this requires a form of memory which Pylyshyn previously argued separate from early vision (404-405). Tsotsos brings forth the concept of other sensory modalities as penetrators of early vision, causing a cross-modality integration that founds cognition; Pylyshyn is quick to dismiss this as not occurring at the early vision level and, once again, denies the modality aspect (406). Attention, a subject to which he devotes a section later, is a cognitive intervention that direct focus, but it does not change perception within that visual field and is not a top-down process (405). Similarly, he denies attention shifts and anticipation as cognitive penetrators, only directors of focal point and pre-visual effects. Neuroimaging by Rhodes and Kalish seem to imply that activation of the v1 region during vision is indicative of cognitive penetration, but Pylyshyn says, “There remains a big gap between showing that there is activity in a certain area during both vision and mental imagery and showing that the activity represents the influence of cognition on perceptual content” (406).

Where is the boundary of early vision?

Several commentators (e.g. Cavanagh, Bermudez, Gellatly) have alluded that a more clear distinction should be made in the definition of early vision by examining the role of conscious and unconscious cognition. Noe and Thompson in particular believe Pylyshyn is implying that all of vision must be unconscious if it is cognitively impenetrable. However, Pylyshyn believes that what separates early vision from other visual counterparts does not involve a criterion of unconsciousness. He does not commit to any opinion about the relationship between consciousness and vision because there is nothing conclusive about the nature of consciousness. He believes that some neural computations by the brain are available to the subject in the form of consciousness and some are not.

Aloimonos and Fermuller suggested that Pylyshyn did not expand upon early vision enough to define its boundary, such as including “depth, orientation, and aggregation information concerning visible surfaces” (408). Pylyshyn agrees to these suggestions along with others and enunciates that most visual properties, like luminance and texture, that are detected by a sensitive visual system are not available to cognition.

Other commentators like Hollingworth and Henderson, Rosenfeld, and Sanocki provide supporting remarks that the boundary of early vision might actually be occurring at a higher level than Pylyshyn originally proposed. Hollingworth and Henderson propose that object recognition involves the pre-semantic matching of the information from early vision and memory which can occur in early vision. Rosenfeld and Sanocki suggest that early vision may have a built-in memory database already embedded that allows for ease of rapid recognition. Pylyshyn agrees with their idea that the boundary of early vision may be expanded but asserts that he sees no need for the involvement of memory. Instead, he proposes that early vision assigns objects into categories based on similar visual properties, but such categorization is not and does not require cognitive memory because there is no judgment being made about the object itself.

Papathomas in his commentary provides a visual example in which observers can have alternate perceptions of 3D shape representations but only after given additional suggestions, which suggests cognition does influence early vision because one of early vision’s outputs is 3D shape representation. Pylyshyn finds his example interesting but sees it more as cognitive influences reaching the post-perceptual phase of vision. For ambiguous shapes and figures, early vision might provide the possible visual interpretations, but selecting one interpretation over another is beyond the scope of early vision.

Peterson argues that there are many types of cognition, and there is evidence to suggest that some types influence vision, particularly subsets of intention and knowledge. In one experiment, she used stereograms to show that intention affects early vision and not post perceptual decisions or eye movement. Pylyshyn interprets her findings to suggest that focal attention in the form of intention is occurring outside of early vision. Early vision, he reiterates again, is generating alternative interpretations, but it is focal attention that mediates the perception.

The nature and role of attention

The idea of attention seemed to be more of a semantic debate with Pylyshyn’s commentators. Sowden, Cohen and Kubovy, Schyns, and Rhodes and Kalish struck up conversation on the definition and usage of attention. To summarize Pylyshyn’s replies: attention is object-based, does not require representations to define itself, and must be defined simply in order to avoid a cognitive penetration problem.

Pani brings up the question of how spatial phenomena relate to attention, to which Pylyshyn does admit we do not have a complete grasp on focal attention, but given that patterns of perception are predictable, attention can follow suit (410). Modalities of focal vision and attention are not cognitive penetration, as Yen and Chen suggest, since they are nor used in depth until later processing (410), and Egeth’s suggestion of pictures as localization cues is simply that: cues for attention, not cognitive penetration (411).

Comments on the role of natural constraints

Dannemiller and Epstein seem to think that Pylyshyn uses natural constraints to explain away the inverse problem of vision, but the natural constraints he describes cannot explain complex perceptual phenomena. Pylyshyn admits that individual constraints can be overcome but natural constraints are just one solution to the inverse vision problem.

Yu believes that “it is likely that by embodying reliable relationships in the world, perceptual processes can gain greater efficiency with minimal cost,” supporting Pylyshyn’s proposition of natural constraints (401).

Hollingworth and Henderson argue that context effects should be considered as having the same status as natural constraints because contextual environments can also have an effect even if the resulting representation is inaccurate. Pylyshyn disagrees and says that a natural constraint is always held true with some exceptions because a natural constraint is part of the vision system architecture. Context effect, on the other hand, is not structurally sound like natural constraints.

Subsystems of vision, other modules, and other modalities

As Vallortigara pointed out, early vision must consist of independent sub-processes that interact with each other to create a holistic interpretation of visual input. Others like Bowers, Bruce et al. offer examples of how early vision can be cognitively impenetrable and also interact with other impenetrable systems, such as language and facial recognition.

Gentaz and Rossetti take a different stance by relating Pylyshyn’s theory of discontinuity between vision and cognition with haptic perception. They assert that the cognitive impenetrability of haptic perception suggests a possibility that the sensory systems in general are all impenetrable. Pylyshyn takes issue with their argument for several reasons. 1) He does not see how his thesis is related to their observation because it seems to him that the manual moving of one’s hands across surfaces and objects is an example of cognitive effect. 2) He did not claim all sensory systems are impenetrable. 3) Haptics does not appear to be a single perceptual modality.

Conclusions

Pylyshyn does not refute that cognition influences vision at specific stages for perception to arise, only that a portion of the vision system known as early vision is independent of aspects of cognition. He ends by saying that what he proposed is a working hypothesis but by no means a final analysis on the subject matter and only further research will be able to elucidate the relationship between vision and cognition.

Questions

1. If early vision can form basic beliefs, is it incorrigibly or prima facie justified? How do visual disorders factor into justification?
2. Pylyshyn shies away from the topic of color in early vision, and instead takes on color memory. Does it seem plausible for early vision to contain more than geometric capacities, but color capacities, as well? What would this mean for cognitive penetration?  
3. The relationship between early vision and memory is a point of contention in the commentary and Pylyshyn’s response. Taking into consideration what he says are the outputs of early vision, is memory isolated from early vision?
4. Does Pylyshyn’s discontinuity theory between vision and cognition give grounds to naturalize epistemology? If so, in what way (e.g. replace, transform, separate)?

“The question of why we see things the way we do in large measure still eludes us” (p.341). This article explores the elusive relationship between vision and cognition. Is what we see solely a function of the stimulation we receive on our retina, or is it “cognitively penetrable” and influenced by what we “expect” to see? Furthermore, does what we see change our beliefs and representations of the world? Pylyshyn’s hypothesis is that early vision encompasses the stage of vision in which computational and top-down processing produces a 3D-image, but this stage is impenetrable to cognitive influence. Rather, cognition is constrained to two points: 1) pre-perception allocation of attention and 2) post-perception pattern-recognition decisions.

Questioning the Continuity Thesis

Pylyshyn puts forward four reasons for questioning whether vision is continuous with cognition. First, he argues that perception is resistant to rational influence. To support this claim, he cites optical illusions – even when you “know” two lines are the same length, you perceive them as uneven. Second, he explains that the principles of perception differ from principles of inference, which follow rational rules of reasoning. Principles of perception are responsive only to visually presented information; they do not reflect simplicity. For example, even if parts of an image are blocked, we can picture what is behind the blocked areas quite accurately. Therefore, these principles are “insensitive to knowledge, expectations, and even to the effects of learning” (p.345). Third, Pylyshyn cites clinical evidence from neuroscience that suggests at least partial dissociation of cognitive and visual functions. Finally, Pylyshyn presents methodological arguments that acknowledge the observed effects of expectations, beliefs, and contextual cues, but designates these influences to stages of processing that lie outside of what is called “early vision” (pg. 345).

Arguments For/Against Continuity

Pylyshyn explores three fields which have also explored the continuity thesis. First, he considers progress in artificial intelligence, in which the goal is to design a system that can “see.” The major progress that has come out of this field has been the development of a knowledge-based/model-based systems approach. These systems use stored general knowledge regarding objects to help determine whether said object appears in the field of view. Though this supports the continuity thesis, Pylyshyn argues that in addition to these systems, there needs to be development of systems that contain constraints on interpretation, which are congruent with the impenetrability thesis.

Next, he introduces discoveries in neuroscience indicating that attention can “sensitize or gate the visual field” (p. 347). This research partially favors the continuity theory, as there is evidence for top-down effects that modulate attention. However, there have been no discoveries of cells that influence the interpretation or emotional part of vision. Pylyshyn argues that what they illustrate is a pattern/motion response, not a content (cognitive) response.

Finally, Pylyshyn offers examples from clinical neurology in which pathology, specifically visual agnosia, strongly indicates a dissociation of vision and cognition. Upon impairment of one of the systems, the other continues to function. This view contradicts the continuity theory. Pylyshyn concurs with this evidence, although admits that these observations could certainly be correlational, not causal.

Determining Visual Stages: Methodological Issues

Pylyshyn acknowledges that all evidence provided above has been highly debated due to the complexity of the visual process. He highlights the importance of clarifying the phases of vision to better interpret these and other findings. He reviews methodological problems associated with distinguishing the stages of the visual process. The signal detection theory explores the ways in which humans make decisions about and respond to stimuli. Two phases of vision have been defined: a “perceptual” phase recognizes a stimulus (cognitively impenetrable), and a “decision” phase formulates a response (cognitively penetrable). The former is represented by the sensitivity parameter (d’), which represents the statistical relationship between the presence of the tone and a person experiencing the tone. The latter is represented by the response bias/criterion measure (ẞ), which represents the statistical relationship between the presence of the tone and the formulation of a response. While this interpretation locates such effects in a post-perceptual stage, Pylyshyn argues that this decomposition is generally too coarse to accurately establish when cognitive influences play a role.

It is important to note that the way we usually determine detection of a signal (sensitivity) is through a response (criterion measure), and therefore, being able to determine the location of cognition fails here as well. However, one way to compensate for this is through measurement of event-related potentials, which allows for measurement of stimulus evaluation uncontaminated by the response decision-making process. Pylyshyn also finds fault with this method because it encompasses everything except the response selection process, including memory retrieval for recognition, decisions, and inferences. He argues that “we need to make further distinctions within the stimulus evaluation stage so as to separate functions such as categorization and identification, which require accessing memory and making judgments, from functions that do not” (p. 351).

He concludes this section by further questioning the relationship between the stages of perception and sensitivity, which seems to be rather inconclusive. It is obvious that we need to determine a sort of mechanism that can lead to specificity in our sensitivity. Therefore, we need some sort of filtering to formulate the hypothesis generation stage of visual perception.

Constraints and Attention

Pylyshyn discusses several examples in which vision “appears on the surface to be remarkably like cases of inference” (p. 354). In these cases the visual system appears to “choose” one interpretation over other possible ones, and the choice appears remarkably “rational” (p. 354). However, Pylyshyn insists the examples do not actually constitute cognitive penetration for two reasons. First, he argues against cognitive penetration of natural constraints, which “are typically stated as if they were assumptions about the physical world” (p. 354). Natural constraints fail to demonstrate inference because the visual system evolved to work as it does, and principles of the visual system are internal to the system. They are neither sensitive to beliefs and knowledge, nor accessible to the cognitive network (p. 355). Next, Pylyshyn argues that cases of so-called perceptual “intelligence” and “problem-solving” also fail, and for the same reasons as natural constraints (also p.355). Specifically, the visual system often fails a simple test of rationality when certain basic facts about the world known to every observer.

How Knowledge Affects Perception

Pylyshyn discusses apparent counterexamples to the discontinuity theory. While he acknowledges the value of these as having an impact on response time, he asserts that the improved response is the result of knowing where to look or what to look for, limiting cognition to the pre-perceptual and post-perceptual stages. Hints of finding meaningful images do not actually help and does not affect the content that is seen (which is required for cognitive penetration). “Expert perceivers” (p. 358) appear to have knowledge that increases their ability to perceive certain patterns and with increased speed, but this seems to result solely from learned classification of visual patterns to enhance recognition and identification. He argues that this is part of the post-perceptual process. Additionally, although findings show that what people see is altered through experience, he argues that cognition plays a role only in pre-early vision by indexing spatially relevant locations. This development of focal attention is an important mechanism by which vision is malleable to the transient external world, and represents the main interface between vision and cognition. Further, he allows that cognition can influence post-perception decision making based on knowledge and experience (although with practice, this can become automated and cognitively impenetrable).

Output of Visual Systems

Pylyshyn makes the case for the visual system being a single system with two outputs, neither of which is knowledge-dependent. 

He defines “early vision” from a functional perspective, as the “attentionally modulated activity of the eyes” (p. 361). In this way, he’s acknowledging that early vision happens after attentional gating, and depends on inputs from other modalities, including non-retinal spatial information. Pylyshyn presents research showing that the output of the visual system in categories, such as shape-classes. Pylyshyn outlines that we form a 3D representation of surfaces (independent of knowledge), encode the layout of a scene (again, without knowledge or reasoning), and perceive a set of surfaces in depth. However, he stresses that “computing what the stimulus before you looks like…does not itself depend upon knowledge” (p.361).  Identifying a shape is not the same as recognizing what it actually is. For that, we need to draw upon other knowledge in memory and perform top-down processing, which can form a bridge between seeing the shape and knowing what it represents.

A second type of visual output is that which affects motor actions. These, too, Pylyshyn argues, happen separately from cognition, and he uses examples from clinical neurology to illustrate the “fractionation of the output of vision” (p.362) that allows patients to act as if they can see even when they don’t “think” that they can.

Questions

• If vision is not continuous with cognition, does that suggest that perception cannot constitute basic beliefs, thus undermining Foundations Theory?
• Is Pylyshyn’s hypothesis consistent with Traditional Epistemology or Natural Epistemology?
• Given the information put forth in this article, as scientists, do you think that we can accurately “observe” any visual data?
• Do you believe that there are two separate forms of visual output? If so, where do you think the motor-function output is derived from?

Patricia Churchland — Epistemology in the Age of Neuroscience

Churchland presents the idea that the framework in which we consider how humans perceive, learn, and understand must be re-contextualized in the age of neuroscience (Figure 1; Churchland p. 545). This new framework calls for the alteration of the question how is it possible for us to represent reality to how does the brain work, by her assertion, and she details three factors to elucidate why contemplating naturalizing epistemology is relevant now. The author postulates that 1) technological developments within the last twenty years have allowed for an understanding of neural systems to such a degree that theorizing about how macro effects, such as human behavior, stem from our neural underpinnings; 2) inexpensive and accessible computing technology allows for investigation of properties at a circuit level (i.e., neural nets); and 3) clinical neurology, biology and psychology are continuously yielding information about the capacity of neural systems, which allows for a segue into considering contemplation of brain, learning, and consciousness with respect to our understanding of evolution. Evolution, according to the author, is critical to consider for three reasons; 1) as the human brain is similar in structure, organization, and componentry to other primate brains, we must anticipate that how we learn, see, hear, and remember does not fundamentally diverge from these processes in other organisms; 2) cognition cannot be removed and considered separately from an organism’s niche and evolutionary goals, and exists in relation to the organism’s necessity to feed, fight, flee, and mate. It therefore cannot be considered in isolation, as though existing for its own sake, and sensorimotor advancements are in place to enhance an organism’s chance of survival; and finally, 3) as the human brain is a product of evolution, which works by modifying existing structures rather than by rewriting the basics, we cannot expect engineering perfection in its design. Churchland then presents a connectionist model as an alternative to the “grand old paradigm” in order to explain our cognitive processes. Connectionist models are constructed to mimic the biology of neural networks, with processing units (neurons), connections between these units (axons and dendrites), and relative weights (strength of synapses). The complexity and dynamism possible in these models, such as the NETtalk model that converts written text to speech, indicate that it may be possible to one day model exactly how our representations are formed and how we reason on a neurobiological level, namely through distinct patterns of activation. Churchland concludes by indicating that progress in this respect will require the efforts of researchers from a variety of disciplines.

Robert Almeder — On Naturalizing Epistemology

In this paper, Robert Almeder grapples with Quine’s argument for naturalized epistemology and concludes that there is ultimately no sound argument in favor of his position. On a Quinean account, naturalized epistemology is a branch of natural science and the only legitimate questions about human knowledge are those answerable through natural science. This version of naturalized epistemology therefore supplants “first philosophy” traditional epistemology—that we can have a notion of whether or not our scientific knowledge satisfies a philosophical analysis of justification or knowledge. Accordingly, we only have scientific knowledge and knowledge of the mechanisms productive of successful practice of natural science.

 The most interesting question to keep in mind during the article is raised by Almeder in his conclusion: “is there something fundamentally incoherent about arguing philosophically for naturalized epistemology… that denies philosophical arguments will count when it comes to answering questions about the nature of epistemology (278).”

Quine’s Argument 

According to Quine, traditional epistemology is concerned with showing how the foundations of knowledge—natural science or mathematics—reduce to certainty. Quine essentially understands traditional epistemology to be the pursuit of refuting the Cartesian sceptic. However, Quine believes traditional epistemology has failed its fundamental pursuit. As examples he points to the reduction of mathematics to the axioms of set theory which do not show how mathematical certainty is possible. Furthermore, Quine points to the Humean issue of induction which holds that no foundation of certainty can be associated with sense impressions or physical objects (natural science). Therefore traditional epistemology is dead and all questions and doubts are scientific in nature, answered by scientists through the natural sciences.

Response to Quine’s Argument

In this section Almeder raises objections by Stroud and Sosa and raises three issues with Quine’s argument. First, Quine is accused of incoherency by employing a philosophical argument to argue against the idea of traditional epistemology. Secondly, a further incoherency is found in the fact that the thesis for naturalized epistemology rests on two premises that are only sound if philosophical arguments about human knowledge are sound. These premises are that 1. Hume’s skepticism is established and 2. There is no analytic/synthetic distinction. Finally, Quine’s argument is accused of mischaracterizing traditional epistemology as being primarily concerned with establishing certainty when it is just as concerned with defining concepts of knowledge. i.e. what it means to know.

Philosophy is Science Argument

This section presents Lycan’s defense of Quine’s argument. Lycan, unlike Quine, believes in the analytic (apriori)/synthetic (empirical) distinction—therefore he does not adopt the Humean argument and face the same criticism as Quine. By Lycan’s account, classical philosophy assumes a deductive model in which indisputable truths are arrived at by deductive reasoning. However, he claims the premises for the deductively reached conclusions must be rooted in non-deductively gained, putative knowledge. Therefore premises are evaluated on their plausibility much like scientific theories. According to Lycan this means that 1. philosophy is in fact a high-level science and 2. proper philosophical methods cannot differ from scientific methods.

Ultimately, Almeder argues that Lycan’s position is untenable for a number of reasons.

Most significantly, philosophy is very different from science in that “in science, but not in philosophy, a necessary condition for any explanation being even remotely plausible is that it be in principle empirically testable” (270).

Traditional Epistemology Will Become Irrelevant Argument

Next, Almeder considers arguments in favor of naturalized epistemology from the eliminitivist position which claims that traditional philosophy will be revealed to be “folk psychology” and eliminated in favor of future neuroscientific accounts of cognitive functioning. Such arguments are favored by the Churchlands and Giere. Almeder believes that while the argument is possible there is really no way to justify such optimism a new model of inquiry. Later he offers objections to Giere, but given the weakness of Gyre’s argument, this section is relatively insignificant.

The Argument From Evolutionary Theory

Here, Lycan addresses two arguments that claim the only valid questions about the nature of human knowledge are those that can be answered in biological science by appeal to evolutionary theory. However, only Kornblith’s argument is even evolutionary in nature and her argument is readily dismissed by Lycan for its false premise P2, which claims that humans have a predisposition for believing truths.

The Impossibility of Defining Justification Argument

Lycan considers Richard Ketchum’s argument to be the most interesting argument in favor of naturalized epistemology. Essentially, Ketchum argues that traditional epistemology must include an acceptable definition of the concept of justification. However, given that the question, “are you justified in accepting this definition of justification” can only be answered in terms of the acceptable definition, there is no non-question begging way of defining the concept justification.

The best response Lycan musters is to say that the question “are you justified in accepting this definition of justification,” is in itself meaningless. This is the case because if there is no definition of justification in mind when the question is asked, then it cannot be known what counts as an answer to the question. Given Ketchum’s argument, however, there is no non-question begging way for a definition of justification to be given.

Questions:

As I asked at the beginning of the summary and as Lycan wonders in his conclusion, do we think that it is fundamentally incoherent to give a philosophical argument in favor of naturalized epistemology?

Does Lycan’s final rejection of Ketchum’s argument that I summarized seem strange at all? Is there a way to defend Ketchum’s argument against Lycan’s critique?

This is a 300-level philosophy course. Consequently, active, student-led discussion is expected. I have found the following format works well for achieving this goal: Each student, hereon called a presenter, will be responsible for leading at least one class (Some of you will do this as teams of two). In effect, these will be no different than days in which I’m presenting, save for the addition of 3 distinct stages leading up to the class:

Stage 1: Summary papers: Each presenter should write a 1000-1500 word paper. See below for details.

Stage 2: Pre-class questions for presenters: Approximately five students who are NOT PRESENTING will be “ON CALL.” Being on call means having to post a question on the course blog in response to the Summary Paper and readings. See details below.

Stage 3: Discussion: Each class will begin with the presenter providing a brief (5-minute) synopsis of the reading. (Everybody will have already read the primary material and the summary paper.) The presentation must conclude with a summary of the main questions that people asked (presenters should group questions according to similar concerns). Students who are neither presenters nor on-call are expected to be the first people to answer these questions. This way, everybody has an opportunity to be involved in each discussion in each class.

Presenters and on-call members who submit everything on time and exhibit an honest effort get at least a 92.5 on these assigments. Those who do exceptional work earn 100. Those who fail to be timely or who fail to exhibit of an honest effort (as judged by me) will earn lower grades.

Stage 1: Summary Paper

As noted above (Stage 1), presenters must write a paper in which:

• The presenter summarizes the texts to be discussed for the class.
• The presenter critically engages the texts he/she summarizes so as to stimulate discussion by, e.g., raising potential objections, exploring potential answers to those objections, etc.

This paper should be posted to the blog no later than 11:59pm on two days BEFORE the presentations (i.e. either Sunday or Tuesday, depending on when you’re presenting). The blog address is:

https://sites.middlebury.edu/ratcog/

Please enter this as a new post. Since there are two sections of this course, this will be a cooperative project with your counterpart in the other section.

Stage 2: On-Call Responsibilities

For any given class, approximately 4-5 students who are not presenting will be on-call. Possible questions might be about:

• Passages (in either the text or the presenter’s paper) that were not clear to you. Cite page numbers and use direct quotations
• Passages (in either the text or the presenter’s paper) that you strongly disagree with. Cite page numbers and use direct quotations.
• For those of you who post a bit later than your peers, you’re encouraged to use other on-call group members’ questions and ideas as the basis of your own question. If you do so, try to tie it back to the readings.

You should post these questions to our blog:

https://sites.middlebury.edu/ratcog/

These should posted as replies to the summary paper. You should post these questions by 2pm on the day before the presentation (so either Monday or Wednesday, depending on when you’re on call.) Please post these as replies to the presenter’s blog entry. This blog is not publicly accessible (i.e. people can’t find it if they Google), so don’t be shy about asking a question that you think is “dumb.” (Chances are, it’s not.)

EVERY student has the following responsibilities for EVERY class:

• To read EVERY question on the blog.
• To think about potential responses to 3-4 questions, and be willing to share your ideas in class.

Stage 3: Presentation/Discussion

On the day of the class in which you are scheduled to present, you will lead discussion. The major questions we will answer are those asked by the On-Call group. Wherever possible, students who are neither presenting nor on call should make the first attempt to answer these questions. (This way, everyone’s involved.) Having said this, it’s sometimes easier to start conversations by asking on-call members to motivate/clarify their questions. Presenters should keep the following in mind:

1. All presenters must use PowerPoint, Beamer, or Keynote.
2. Slideshows must be less than 10 slides long.
3. The presenter should begin by situating the reading within the broader themes of the course, especially with respect to earlier readings.
4. The chief criterion by which presenters will be assessed is how well they stimulate discussion. (This is not as easy as it looks!) Consider how different questions on the blog hang together; find interesting points of disagreement.

The slideshow should not be overflowing with information, yet must be sufficiently clear that other members of the class find the ideas easy to follow. Consequently, presenters must be especially reflective about how the accompanying commentary will supplement what the slides say. You are encouraged to look for additional materials about how to use PowerPoint effectively in academic presentations.

Philosophers (and others) study how we ought to reason. By contrast, psychologists (and others) study how actually do reason. Often, their findings conflict. How should these conflicts be reconciled? Potential topics include different kinds of reasoning (deductive, probabilistic, explanatory, analogical, practical, etc.), naturalized epistemology, theories of justification, and heuristics and biases. Prerequisites: either PHIL0180 or PSYC0105.