P303: BH PART A

What is the lecture's "causes versus reasons" thought experiment?

Natural sciences explain events by causes that operate the same way every time. The brain is a physical organ obeying natural laws, yet we explain behaviour by reasons: beliefs, desires, intentions, meanings, expectations. Can reason-explanations be grounded in physical causes? KEY POINTS AND SOURCES: Sets up the mind-body problem that bookends Part A (introduction and Section 3). Cognitive science's answer: information is the bridge, since a belief is a physical pattern (a cause) that carries content (a reason). (ext.) Davidson (1963): reasons are a species of cause.

State the mind-body problem as the lecture frames it.

Is the mind explained by the same principles as the physical world, or by non-physical entities? If thoughts cause actions, are mental and physical the same thing, or can a mental event cause a mechanical one (and vice versa)? KEY POINTS AND SOURCES: Relevant to perception (stimulation produces mental content) and to motor control (intentions produce movement). Answer: mental states are physically realised information.

Reason 1 for cognitive science?

The mind is not a blank slate. We are born with sensorimotor reflexes, working sensory transducers, and the capacity to develop and self-teach them into intelligent systems. KEY POINTS AND SOURCES: Few theorists defend a pure blank slate. Links forward to the visual cliff (early depth perception) and innate constraints.

Which innate core cognitive systems are proposed?

Universal grammar; inanimate objects and their mechanical interactions; agents and goal-directed actions; sets and numerical relations (ordering, addition, subtraction); places and their geometric relations. KEY POINTS AND SOURCES: Universal grammar: Chomsky. (ext.) Objects, agents, number and geometry match Spelke and Kinzler's (2007) core knowledge systems.

Reason 2 for cognitive science?

The mental world is grounded in the physical world by information, computation and feedback. KEY POINTS AND SOURCES: Begins to answer causes versus reasons. Note: this is a disembodied, information-processing picture, which Section 2 challenges.

Map perception, beliefs, thinking and acting onto information terms (Reason 2).

Perception = extracting information from sensory stimulation. Beliefs and memories = stored patterns of information. Thinking and planning = transformations of those patterns. Doing and acting = overt behaviour informed by predictions and corrected by feedback. KEY POINTS AND SOURCES: These four map onto the lecture's structure: perception (S1), stored knowledge shaping perception (S2), action with prediction and feedback (S3).

Reason 3 for cognitive science, with examples.

Infinite behaviour from finite combinatorial programs. DNA: four bases build 20,000 to 25,000 genes and endless human variety. Language: finite phonemes and rules yield infinite sentences. Chess: fixed rules, virtually infinite games. Music: finite notes and rhythms, endless melodies. KEY POINTS AND SOURCES: Generativity is a hallmark of mind; motor equivalence in S3 (one plan, many executions) is a related idea.

Reason 4 for cognitive science?

Universal cognitive mechanisms underlie superficial cultural variation. Customs and languages are learned, but the mechanisms for learning are innate; deep cognitive structure is more evolutionarily stable than surface behaviours. KEY POINTS AND SOURCES: Humans share deep cognitive structure even when their languages are mutually unintelligible.

Whose account do the four reasons come from?

Steven Pinker, The Blank Slate (2002). KEY POINTS AND SOURCES: (ext.) Not named on the slides, but an accurate attribution.

Name the five sections of the lecture and which are in Part A.

Part A: Introduction (why perceiving and doing are thinking); S1 perception as sensorimotor exploration, Umwelt, four families of theory; S2 embodied cognition; S3 cognitive science of motor control. Part B: S4 active vision and scene perception; S5 reading print and braille. KEY POINTS AND SOURCES: The unifying claim of Part A: perceiving and acting are forms of thinking.

What is the logic of the "drawing what you see" exercise?

IF you see perfectly AND your pencil control is skilled, THEN your drawing should be a hand-drawn photograph. It is not, so something failed: perception (distorted percept), decision (suboptimal choices about what to draw), or execution (inadequate movements), or other factors. KEY POINTS AND SOURCES: The three links preview the lecture: perception (S1), cognition shaping perception (S2), motor control (S3).

Why is a drawing from life rarely like a photograph? (perceptual answer)

Perception delivers the distal object with object constancy (stable size, shape, colour), not the proximal retinal image. To draw realistically you must reproduce the proximal image, so you must undo constancy. Knowledge of the real shape (for example, a rectangular table top) leaks into the drawing. KEY POINTS AND SOURCES: Integrative answer also cites S2 (seeing x as x; allocentric ventral coding) and S3 (closed-loop skill, speed-accuracy, coordinate mapping).

Give the perceptual chain from world to experience.

Distal object (the thing in the world) → proximal stimulus (the retinal image) → perceptual object (what is experienced, ideally matching the distal object via constancy). KEY POINTS AND SOURCES: Lecture slide heading: "From distal object to perceptual object; object constancy".

What is object constancy?

Perceiving an object's real properties (size, shape, colour) as stable despite large changes in its retinal image with distance, angle and lighting. KEY POINTS AND SOURCES: Constancy is evidence that vision solves the inverse projection problem in favour of the distal object.

Define Umwelt. Who introduced it?

Jakob von Uexküll ("fon OYKS-kül"). German for "environment": the complex but limited, species-specific sensory world an animal is sensitive to. Animals in the same ecosystem pick up different signals and different ranges of them. KEY POINTS AND SOURCES: Core point: perception is a filter shaped by a species' ecological needs, not a camera recording all of reality.

Give animal examples of different Umwelten.

Dog: strongly olfactory (some colour vision, excellent hearing). Tick: temperature and odour. Echolocating bat: dominated by air compression waves. KEY POINTS AND SOURCES: Same environment, different worlds. Use two contrasting examples in an answer.

How much of the electromagnetic spectrum do humans sense, and what follows?

About one ten-trillionth. Yet for us the visible spectrum is "All. There. Is." We do not notice what we miss. KEY POINTS AND SOURCES: We treat our Umwelt as the equivalent of all there is and accept reality as presented to our senses.

What does the lecture say about blind people and Umwelt?

Congenitally blind people have a different Umwelt and rarely complain of lacking vision: you do not miss what you never had. The lecture raises the harder case of losing sight after 20 or 30 years. KEY POINTS AND SOURCES: Electromagnetic radiation is simply not, or no longer, part of their Umwelt.

What do Groucho Marx's line and the nature-nurture questions introduce?

"Who are you going to believe, me or your eyes?" captures the tension between knowledge and perception. Also: are we born with sensory systems that must learn to perceive, or with ready perceptual machinery needing only exposure? KEY POINTS AND SOURCES: Tension between knowledge and seeing runs through S1 (inference vs constraints) and S2 (cognitive penetrability). The visual cliff bears on nature-nurture.

What is the guiding question of visual perception theories, and why is it urgent?

"How do we see the world the way we do?" (Köhler, per the lecture). Urgent because, for most theorists, the mind solves an impossible problem: the inverse projection problem. KEY POINTS AND SOURCES: Note: the closely matching Gestalt question "Why do things look as they do?" is usually credited to Koffka (1935) (ext.).

Define the projection problem.

Given an object's real size (h) and egocentric distance (d), projective geometry specifies a single, unique retinal image (visual angle α). Given d and h, compute α. Not trivial, but it has one solution. KEY POINTS AND SOURCES: α = 2 arctan(h / 2d). Well-posed.

Give three applied examples of the projection problem from the lecture.

Film set design (if the camera is here with this field of view, how must the set be built to look real?). Julian Beever's pavement art (correct from one viewpoint only). NASA New Horizons Pluto image from 77,000 km, claiming 3,500 m mountains. KEY POINTS AND SOURCES: Pluto question "calculations or estimations?": scale must come from added assumptions (sun angle, shadows), just as the visual system must add assumptions.

Define the inverse projection problem.

Vision must recover the distal layout (d and/or h) from the proximal image (α) alone, with no independent information about distance or size. Given α, compute d and h. KEY POINTS AND SOURCES: The most examinable concept in S1.

Why is the inverse projection problem ill-posed?

Infinitely many distal layouts produce the same retinal image (a small near object or a large far one), so there is more than one possible solution. KEY POINTS AND SOURCES: Ill-posed = ill-defined = no unique solution. Same logical form as the degrees of freedom problem in S3.

What two things must every theory of vision explain about the inverse projection problem?

(1) How the visual system solves a problem with no single solution, quickly and for novel scenes. (2) Why its solution is so frequently correct. KEY POINTS AND SOURCES: Each of the four families answers differently.

What happened at Erebus and what does it illustrate?

28 November 1979, Air New Zealand Flight TE901, an Antarctic sightseeing flight, crashed into Mount Erebus, Ross Island. All 257 on board died; NZ's worst peacetime disaster. KEY POINTS AND SOURCES: Illustrates catastrophic failure of the inverse projection solution when cues are absent. (ext.) Altered flight path coordinates plus sector whiteout removed texture, shading and horizon, so a rising white slope looked like flat terrain.

Name the four families of theories and their key theorists.

Constructivism (Helmholtz, Rock, Bruner); Computationalism (Marr, Biederman); Predictive models (Bayes, Grush, Clark, Friston); Direct realism (Gibson, Warren, Chemero). KEY POINTS AND SOURCES: Hook C-C-P-D.

How does each family deal with the inverse projection problem?

Constructivism: cues to depth plus prior knowledge. Computationalism: constraints on solutions. Predictive: Bayesian inference, continually updated. Direct realism: a pseudo-problem, solved by optic flows and affordances. KEY POINTS AND SOURCES: One-liners: indirect inference; computation over representations of images; predictions plus confidence in input; direct and non-computational.

Constructivism's core claim in one line?

Perception is indirect inference: it is constructed from bottom-up sensory cues and top-down knowledge. KEY POINTS AND SOURCES: Solves the inverse projection problem with cues to depth plus prior knowledge.

Why are both bottom-up and top-down information necessary for constructivists?

Without bottom-up input, perceptions would be hallucinations detached from the world. Without top-down input, perceptual systems would work endlessly to solve an impossible problem. KEY POINTS AND SOURCES: Neither is sufficient; both are necessary.

What is Helmholtz's unconscious inference?

Fast, automatic inference (the lecture links it to "System 1") that uses bottom-up and top-down sources to produce a single, usually accurate percept. KEY POINTS AND SOURCES: Helmholtz (1867). Bayes later formalises it (predictive models).

How do constructivists characterise the perceptual process?

Neither single-direction nor single-step; it needs extensive experience and intelligence. It is subconscious, intelligent problem-solving by fast processes, immune from conscious beliefs, involving: drawing inferences; drawing on past experience; basing one perception on another. KEY POINTS AND SOURCES: Example of basing one perception on another: size judged from perceived distance.

Why, according to constructivism, are these processes subconscious?

Precisely because they are so important: it takes strenuous reflection to notice which cues we use, and we are rarely aware of the past experiences behind a given interpretation. KEY POINTS AND SOURCES: Explains why perception feels effortless and direct even if it is inferential.

What did Rock and Bruner contribute?

Irvin Rock: perception as problem-solving (The Logic of Perception, 1983). Jerome Bruner: the "New Look", showing expectations, needs and values shape perception. KEY POINTS AND SOURCES: (ext.) dates and titles; Bruner (1957) "On perceptual readiness".

List the monocular depth cues and what each signals.

Occlusion (covering object nearer); relative size (smaller image farther); familiar size; height in visual field (nearer horizon = farther); linear perspective (parallel lines converge); texture gradient (finer = farther); atmospheric perspective (hazier = farther); shading and shadows; motion parallax (near objects sweep past faster). KEY POINTS AND SOURCES: For constructivists these are ambiguous cues combined with knowledge; for Gibson, texture gradients and motion are information, not cues.

Describe the Ames room.

A trapezoidal room with one far corner much more distant than the other. Seen monocularly through a peephole it projects the same retinal image as a rectangular room, so equal-height people in the two corners look very different in size, and a person walking across seems to grow or shrink. KEY POINTS AND SOURCES: (ext.) Adelbert Ames Jr., 1940s.

Why is the Ames room constructivism's star demonstration?

The image fits infinitely many rooms; vision picks the rectangular one because prior experience says rooms are rectangular. That prior is strong enough to override familiar size (people do not shrink). KEY POINTS AND SOURCES: Illustrates the inverse projection problem solved by knowledge. Gibsonian reply: move your head and the illusion collapses.

How does Julian Beever's sidewalk art relate to the Ames room?

It is inspired by the same logic: an anamorphic drawing built so its projection from one viewpoint matches a 3D scene. From any other viewpoint it is distorted. KEY POINTS AND SOURCES: Exploits the projection problem and our prior assumptions.

Strengths of constructivism?

Explains illusions elegantly (normally valid assumptions applied in abnormal conditions); explains the role of learning, context and expectation; explains why perception is rapid yet sophisticated. KEY POINTS AND SOURCES: Best fits vision-for-recognition (ventral stream).

Weaknesses of constructivism?

(1) If perception is inference, why do illusions persist when we know they are illusions? (2) Relies on impoverished, static, monocular stimuli (peepholes) that Gibson calls ecologically invalid. (3) The visual cliff suggests depth perception needs little experience. KEY POINTS AND SOURCES: Weakness 1 is the computationalist objection: unrevisable "inference" looks more like a constraint.

Computationalism's core claim?

Perception involves computations over representations of images; the inverse projection problem is solved by constraints on possible solutions. KEY POINTS AND SOURCES: Key names: David Marr, Irving Biederman.

Quote Marr's definition of vision.

"Vision is a process that produces from images of the external world a description that is useful to the viewer and not cluttered with irrelevant information." KEY POINTS AND SOURCES: Marr (1980, p. 31), as cited. Also: the task is "to reliably derive properties of the world from [retinal] images of it" (p. 23). The word "useful" returns in S2: perception must speak the language of thought and action.

Name Marr's three levels of analysis.

Level 1: computational theory. Level 2: representation and algorithm. Level 3: hardware implementation. KEY POINTS AND SOURCES: Marr's levels are a general framework you can apply to any cognitive system, including motor control.

What questions does Marr's Level 1 ask?

What is the goal of the (visual) system? By what process is the retinal image used to realise the goal? What is the logic of the strategy? KEY POINTS AND SOURCES: The "what" and "why".

What questions does Marr's Level 2 ask?

How can the computational theory be implemented? What are the input and output representations? What algorithm transforms input into output? KEY POINTS AND SOURCES: The "how".

What questions does Marr's Level 3 ask?

How are the representations and algorithms physically realised? How many realisations are possible? Which one do humans (or fish, bees, birds) have? KEY POINTS AND SOURCES: The "where" and "in what".

Which of Marr's levels is most important psychologically, and why?

Level 1. It matters more to know what is computed and why than where in the brain or by which algorithm; answering the computational question helps constrain the other two. KEY POINTS AND SOURCES: Analogy: you cannot understand flight from feathers alone; you need aerodynamics.

What are Marr's computational stages of vision?

Retinal image → primal sketch (edges, blobs, bars) → 2½-D sketch (viewer-centred surfaces and depth) → 3-D model (object-centred representation). KEY POINTS AND SOURCES: (ext.) Marr (1982), Vision. The primal sketch reappears in S2 as an "early", cognitively impenetrable process.

How are objects represented with generalised cones?

A simple object is one generalised cone (a cross-section swept along an axis); complex objects are combinations of cones in a nested hierarchy, each with its own cone and axis (body, then limbs, then fingers). KEY POINTS AND SOURCES: (ext.) Marr and Nishihara (1978).

What are Biederman's nonaccidental properties?

Image properties very unlikely to arise by accident of viewpoint (collinearity, curvilinearity, symmetry, parallelism, co-termination), used to identify volumetric primitives (geons) in recognition-by-components. KEY POINTS AND SOURCES: Biederman (1987). They survive viewpoint change, so they are a principled partial solution to the inverse projection problem.

What are perceptual constraints?

Properties of the world that have always been true in the environment where our perceptual systems evolved, operating like hard-wired computational routines that restrict which interpretations are possible. KEY POINTS AND SOURCES: A "recent development" in explaining how the inverse projection problem is solved.

Explain the bumps and craters demonstration.

If the display were truly ambiguous, about 50% would see one half as bumps and 50% as craters. Instead nearly everyone sees the same thing, and flipping the image swaps them. KEY POINTS AND SOURCES: The visual system assumes light comes from above (the sun throughout evolution). (ext.) Ramachandran (1988).

Name other natural constraints proposed by computational theorists.

Matter tends to be cohesive and opaque; surfaces are uniformly coloured. (Plus light from above.) KEY POINTS AND SOURCES: Constraints explain why illusions work even when you know they are illusions.

Give three differences between constraints and inferences.

Location: constraints are in hard-wired perceptual circuitry; inferences occur in the language of thought (mentalese). Scope: constraints are modality-specific and inaccessible outside their sensory system; inferences are general and may be amodal. Revisability: constraints operate automatically and are not overridden by beliefs; inferences can be modified by instruction, experience, language or imitation. KEY POINTS AND SOURCES: High-yield comparison table on the slides.

Why can constraints explain the persistence of illusions better than inference?

Constraints cannot be overridden by higher-level knowledge, so knowing the lines are equal does not change the percept. Revisable inferences predict that knowledge should dispel the illusion. KEY POINTS AND SOURCES: Links to informational encapsulation (Fodor) in S2.

Describe the visual cliff and its result.

A glass table with a checkerboard just under the glass on the shallow side and far below on the deep side. Most crawling infants (about 6 to 14 months) cross the shallow side but refuse the deep side, even when their mother calls. KEY POINTS AND SOURCES: (ext.) Gibson and Walk (1960).

What are the implications of the visual cliff?

Depth perception and its link to action appear very early with little visual experience, favouring an innate or rapidly maturing contribution (constraints, or Gibson's directly available information) over purely learned inference. KEY POINTS AND SOURCES: Caveat: crawling infants are not newborns, so some experience, especially self-produced locomotion, cannot be ruled out.

Strengths and weaknesses of computationalism?

Strengths: rigorous and implementable (computer vision, AI); Marr's levels are a general framework; constraints explain persistent illusions. Weaknesses: starts from the static retinal image so accepts the inverse projection problem as real; constraints are often identified post hoc. KEY POINTS AND SOURCES: Gibson: both constructivism and computationalism study a frozen, impoverished observer.

Direct realism's core claim?

Perception is direct and non-computational: the dynamic stimulus field of a moving perceiver contains enough information to specify spatial layout, so the inverse projection problem is a pseudo-problem. KEY POINTS AND SOURCES: Originated with J. J. Gibson. (ext.) The Ecological Approach to Visual Perception (1979).

Why are stimulus fields dynamic, according to Gibson?

Objects are often in motion and, even more reliably, perceivers are always moving. KEY POINTS AND SOURCES: Movement generates the information that resolves ambiguity.

What are Gibson's two claims about cues?

(1) The "cues" to layout are already combined in the flux of light reaching the eyes; they are more than cues, they are information. (2) Perception requires no top-down operations or knowledge-mediation, no inference and no construction of percepts. KEY POINTS AND SOURCES: Direct contrast with constructivism's cues plus knowledge.

What is perception at an "ecological scale"?

Natural perceptual behaviours in natural contexts, at the scale of animal and environment, where perceivers move in order to perceive and perceive in order to move. KEY POINTS AND SOURCES: Direct realists argue these, not lab stimuli, are the appropriate stimuli to study.

Why prefer "information extraction" to "information processing"?

Information is already combined when it reaches the eye; what is already in stimulation need not be computed, added to, filtered, stored, inferred or constructed. It only needs to be picked up. KEY POINTS AND SOURCES: Gibson's "pick-up" of information.

Why does direct realism say "the eye is not a camera"?

Retinal images are not static, and the brain does not have to interpret the ambiguous. Vision is active: eyes, head and body move, and almost always do. Retinal projections are dynamic optic flows specific to (a) environmental structure and (b) the movements we can make in it. KEY POINTS AND SOURCES: The theory is committed to information-in-stimulation and seeks to specify what information is available and extracted.

What do texture gradients provide?

A textured ground whose uniformity produces gradients that specify the layout and supply a scale on which all objects in the scene are anchored. KEY POINTS AND SOURCES: Equal texture units occluded = equal size, so size and distance come "for free". First of three perspective transformations.

What does binocular disparity provide?

The direction and magnitude of disparity vary directly with the distance between perceiver and objects, giving a continuous scale for extracting distance. KEY POINTS AND SOURCES: Second perspective transformation.

What is optic flow and what does it specify?

The perspectives produced over time by moving through a scene. Heading direction is the centre of optic outflow (elements flow outward from it, even if the point itself is not visible). Flow velocity is lawfully related to distance: slow for distant objects, very fast for approaching ones. KEY POINTS AND SOURCES: Third perspective transformation. (ext.) Rate of image expansion gives time-to-contact directly: tau (Lee, 1976).

Define affordance.

The perception of opportunities for action: what a situation permits, allows or provides the opportunity to do. Affordances are not things; they are relations between the structure of the environment and our physical capabilities. KEY POINTS AND SOURCES: Gibson. Seed of embodied cognition and Warren's stair research.

Give the lecture's affordance example and another.

Instead of seeing two panels and a gap, the person sees whether the gap affords passage, for him, and how. Another: a chair affords sitting for an adult but climbing for a toddler. KEY POINTS AND SOURCES: The same environment affords different actions to different bodies.

Why does Gibson think the inverse projection problem is false?

It only arises if the eye is a stationary camera producing a single frozen image. A moving perceiver's transforming optic array (flow, parallax, disparity, texture) uniquely specifies layout, so ambiguity is an artefact of impoverished lab stimuli. KEY POINTS AND SOURCES: Stated test question. Example: the Ames room only works through a static monocular peephole.

How does Gibson's insight reappear in predictive models?

Action is treated as a way of reducing prediction error: instead of updating beliefs, the organism may act so sensory input conforms to predictions. KEY POINTS AND SOURCES: Active inference (Friston). Shows predictive models absorbing Gibson.

Who are Warren and Chemero in this tradition?

William Warren: body-scaled affordances (stair climbing) and optic flow research. Anthony Chemero: radical embodied cognitive science. KEY POINTS AND SOURCES: (ext.) Chemero (2009), Radical Embodied Cognitive Science.

Strengths of direct realism?

Ecologically valid; explains real-time control (driving, catching, landing aircraft); predicts the importance of self-motion; affordances anticipate embodiment. KEY POINTS AND SOURCES: Best fits vision-for-action (dorsal stream).

Weaknesses of direct realism?

Struggles with illusions that persist under natural viewing, with recognising and naming objects by category, with imagery, and with how pick-up is physically achieved (critics say it relocates computation). Says little about seeing x as x. KEY POINTS AND SOURCES: Recognition and meaning are ventral, constructivist territory.

Predictive models' central claim?

Perception is a self-organising (feedback-driven) attempt to minimise the mismatch between (1) internally generated predictions of the causes of sensory input and (2) the sensory events actually experienced. That mismatch is used to modify predictions. KEY POINTS AND SOURCES: A resynthesis of constructivist and computational elements.

What is prediction error?

The mismatch between predicted and actual sensory input; it is the key feature, used to update predictions for subsequent occasions. KEY POINTS AND SOURCES: Also the common currency linking perception and motor control (forward models).

In what sense are predictive models partly bottom-up, and how does this differ from constructivism?

Sensory data are used to test and update predictions, not to construct perception. KEY POINTS AND SOURCES: In constructivism data are building blocks; here they are tests of hypotheses.

In what sense are predictive models partly top-down?

The brain constantly generates probabilistic beliefs about the environment, its structure, and the signals it will encounter. KEY POINTS AND SOURCES: Priors and predictions.

Describe the hierarchy in predictive models.

Higher levels process slower-changing identities (events, objects, intentions); lower levels encode transitory details (edges, tones, movement, colours, textures). KEY POINTS AND SOURCES: Allows refinement at multiple timescales.

Which way do predictions and errors flow?

Predictions flow down the hierarchy; prediction errors flow up. KEY POINTS AND SOURCES: Common exam slip: reversing these.

What is Bayesian inference (in words and formula)?

The probability of a hypothesis given the data is proportional to how well the hypothesis predicts the data (likelihood) times its prior probability. P(H given D) = P(D given H) × P(H) / P(D). KEY POINTS AND SOURCES: Combines prior beliefs with sensory evidence. Formalises Helmholtz's unconscious inference.

Apply Bayes to the inverse projection problem (Ames room).

Many layouts fit the image equally well, so likelihood cannot choose. Priors break the tie: the prior for rectangular rooms is enormous, so the posterior favours "rectangular room, oddly sized people" over "trapezoidal room, normal people". KEY POINTS AND SOURCES: Statistics link: like diagnostic testing, where prevalence (prior) can dominate a positive result (data).

What is precision weighting?

The system estimates the precision (reliability) of sensory signals and predictions. High-precision errors strongly influence perception; low-precision errors are down-weighted. KEY POINTS AND SOURCES: Not all prediction errors are treated equally.

How is attention modelled in predictive models?

As the selective tuning of precision, determining which signals dominate perceptual inference. KEY POINTS AND SOURCES: Unifies attention with perception.

Which three phenomena do the slides group as evidence for predictive models?

Bistability (perception alternates between two equally probable hypotheses, e.g. Necker cube); seeing what is not there (illusory contours); seeing differently when input is identical (context changes priors, not data). KEY POINTS AND SOURCES: Each is a case where priors visibly shape the percept.

Kanizsa triangle, phases 1 to 3.

(1) Bottom-up input is fragmentary: Pac-Man discs and line corners. (2) Higher-level priors activate: closed shapes, symmetry, occlusion, Gestalts. (3) Top-down predictions generate a triangle that, if true, predicts the fragmentary input but includes full edges. KEY POINTS AND SOURCES: The lecture's model example of predictive processing.

Kanizsa triangle, phases 4 and 5.

(4) Prediction error is minimised: the triangle reduces the mismatch between expected input (occluded discs plus triangle edges) and actual input, explaining it better than any alternative, so it becomes the percept. (5) Illusory contours are experienced: the triangle looks brighter than the background and the brain may generate activity for absent contours. KEY POINTS AND SOURCES: The best guess is wrong but still the most probabilistically efficient interpretation.

What does "controlled hallucination" mean?

Perception is the brain's prediction, constrained (controlled) by sensory error signals; illusions are cases where the best guess is wrong. KEY POINTS AND SOURCES: Phrase popularised by Anil Seth (ext.). Contrast with constructivism's warning that no bottom-up = hallucination.

Who are Grush, Clark and Friston?

Rick Grush: emulation theory (internal models spanning perception and motor control). Andy Clark: predictive processing (Surfing Uncertainty). Karl Friston: free-energy principle and active inference. KEY POINTS AND SOURCES: (ext.) Grush (2004); Clark (2013, 2016); Friston (2010).

Strengths of predictive models?

Unify perception, attention (precision), action (active inference) and learning; quantitatively precise; treat illusions as rational outcomes; connect directly to motor forward models in S3. KEY POINTS AND SOURCES: The most integrative of the four families.

Weaknesses of predictive models?

Risk of unfalsifiability (flexible priors and precisions fit almost any result post hoc); "controlled hallucination" sits uneasily with Gibson's ecological evidence; does not itself say where priors come from (innate constraints or learning). KEY POINTS AND SOURCES: Good evaluative point for essays.

Which family treats the inverse projection problem as a pseudo-problem, and which match the two visual streams?

Direct realism denies the problem. Direct realism fits the dorsal (action) stream; constructivism and computationalism fit the ventral (recognition) stream; predictive models integrate both. KEY POINTS AND SOURCES: A+ synthesis: the families answer different questions at different timescales rather than simply competing.

What is Section 2's title question, and its four learning objectives?

"What's inside your head, or what's your head inside of?" Objectives: understand the disembodied vs embodied distinction; complications in understanding illusions; how what you see depends on context; evaluate claims of embodied cognition. KEY POINTS AND SOURCES: Evaluation is an explicit objective, so critique is expected in answers.

Why are seeing x and seeing x as x not the same?

Seeing x registers perceptual properties (shape, colour, location). Seeing x as x categorises or interprets it using stored knowledge, concepts and context (as a cup, as dangerous, as expensive). KEY POINTS AND SOURCES: Stated test question. Where perception hands over to cognition; underlies the perception-versus-judgement critique. Slide examples "Tested" and "Metro": same input, different readings.

Why must perception be compatible with thinking and acting?

If perception's primary purpose is to inform, guide and permit decisions about how to act and move (Marr's "useful to the viewer"), the brain's language of perception must be compatible with its language of thinking and acting. KEY POINTS AND SOURCES: The perception-cognition boundary matters in the real world (eyewitnesses, pilots).

What did the cognitive revolution assume about cognition?

From about 1955 (giving us computers, AI, robotics), cognition was assumed distinct from both perception and action: it depends on input but not on the mode of input, and produces output regardless of the mode of output. KEY POINTS AND SOURCES: People think, but so do computers, sometimes faster and more accurately. What do flesh and silicon share?

List the five features of disembodied cognition.

(1) Thinking is symbol manipulation. (2) Symbols have a physical instantiation (neuronal patterns; binary code). (3) They (temporarily) represent properties of the world. (4) They are manipulated by rules (if-then, and, or, not). (5) They are arbitrarily but consistently related to what they represent. KEY POINTS AND SOURCES: (ext.) Newell and Simon's (1976) physical symbol system hypothesis. Feature 5 is challenged by embodied language research.

Describe the classical three-part architecture of mind.

Sensory input modules and motor output modules (fast, shallow, modular, informationally encapsulated, dedicated neural structures) surround central systems (slow, deep, informationally promiscuous, nonmodular, large neural networks) containing knowledge, beliefs, long-term memory, concepts and categories, language. KEY POINTS AND SOURCES: (ext.) Fodor (1983), The Modularity of Mind.

What do informationally encapsulated and informationally promiscuous mean?

Encapsulated: a module cannot access information held elsewhere (your knowledge cannot change it). Promiscuous: central systems can draw on any information, since anything might be relevant to a belief. KEY POINTS AND SOURCES: Encapsulation explains why the Müller-Lyer illusion persists despite knowing the lines are equal.

What is mentalese?

The amodal language of thought in which central cognition (and inference) is said to occur. KEY POINTS AND SOURCES: Inferences occur in mentalese; constraints do not.