Vision needs non-conceptual connections to objects in the world (just as concepts do)

1. Vision needs non-conceptual connections to objects in the world (just as concepts do) Zenon Pylyshyn Rutgers Center for Cognitive Science • Introduction to a theory of visual indexes (aka FINSTs)

2. Plan of talk: Visual Indexes • Theoretical motivations behind the FINST theory • The need for a primitive mechanism of individuation • Because individuation must be of distal objects, we have the Correspondence Problem: When do two proximal tokens correspond to the same distal object? • A special case: incremental construction of visual representations • Empirical studies of individuation and indexing • Object-specific effects (static & moving objects) • Multiple Object Tracking technique. • What, if any, encoded properties are used to individuate, index, and track objects?

3. Visual Indexes (FINSTs) and what they mean for vision science and cognitive science The need for a mechanism that individuates objects Examples of solving the correspondence problem Individuating distal objects requires solving the correspondence problem Object-based allocation of visual attention A special case of the corre-spondence problem occurs when visual representations are constructed over time. Multiple Object Tracking and Visual Indexes: what it means for connecting vision and the world

4. An important function of early vision is to individuate and select token elements (let’s call them “objects” for now) • The most basic perceptual operation is the individuation and selection that precedes the formulation of perceptual judgments. • Making visual judgments presupposes that the things (objects) that judgments are about have been individuated and selected (or indexed – i.e., made accessible). • Another way to put this is that the arguments of perceptual predicates P(x,y,z,…) must be bound to things in the world in order for the judgment to have perceptual content.

5. Several objects must be picked out at once in relational judgments • For example, when we judge that certain objects are collinear, we must select (and the visual system must be able to refer to) the relevant individual objects.

6. Several objects must be picked out at once in relational judgments • The same is true for other relational predicates, like inside or on-the-same-contour… etc. We must pick out the relevant individual objects first.

7. Several objects must be picked out at once in numerical judgments • In subitizing, the cardinality of sets of 4 or less can be judged rapidly and accurately (over 4 is slow and error-prone). • Subitizing only occurs if items can be automatically individuated.

8. Enumerating different layouts of squaresTrick, L. M., & Pylyshyn, Z. W. (1994). Why are small and large numbers enumerated differently? A limited capacity preattentive stage in vision. Psychological Review, 101(1), 80-102.

9. Another property that cannot be used to subitize: on-same-contour

10. Individuation is different from discrimination

12. How do we select (and index) objects in our field of view? • The principal way we select individual objects is by foveating them – by looking directly at them (Notice that this results in a deictic reference). • We can also select with focal attention, which is independent of direction of gaze. • Focal attention appears to be unitary, yet we can select more than one thing at a time (e.g., in making a relational judgment). So it seems that we need to distinguish attending from selecting: That’s where Visual Indexes or FINSTs come in. • A question for later: In virtue of what properties are primitive objects individuated and indexed?

13. Indexes must individuate and select objects in the world. This leads to the ubiquitous correspondence problem in vision • Apparent motion, stereo vision, tracking, and very many visual computations face the problem of identifying which proximal image-features correspond to the same individual distal object. • Less well known is the correspondence problem faced when a single visual representation is constructed incrementally over time. • The way the correspondence problem is solved determines what the vision system counts as an individual. These primitive individuals (called “objects”) are thus mind-dependent.

14. Example of the correspondence problem for apparent motion The gray disks correspond to the first flash and the black ones to the second flash. Which of the 24 possible matches will the visual system select as the solution to this correspondence problem? What principal does it use? (Dawson & Pylyshyn, 1988) Curved matches Linear matches

16. One of the most troubling forms of the correspondence problem occurs because visual representations are constructed incrementally over time It is clear that when vision requires eye movements, a visual representation is constructed incrementally. But there is also evidence that percepts are built up over time even for the automatic perception of simple forms. So this type of correspondence problem is routine in vision. Why does it constitute a special problem?

17. Example: Drawing a diagram and noticing its properties

18. Some of the distinct “views” while exploring the diagram

19. The correspondence problem for incremental construction of a visual representation • When a property F of some particular individual (token) object O is noticed or encoded, the visual system must check whether object O is already represented. If it is, the new property must be associated with the existing representation of O. • If the only way to identify a particular individual object O is by its description, then the way to solve this correspondence problem is to find an object in memory that bears a particular description (one that had been unique at the time). Which description? If objects can change their properties, we don’t know under what description the object was last stored. Perhaps we look for an object with a description that overlaps the present one, or perhaps we construct a description that somehow incorporates time.

20. The correspondence problem for incremental construction of a visual representation • Even if it were otherwise feasible to solve the correspondence problem by searching for a unique past description, this would in general be computationally intractable (technically, matching descriptions is an NP-hard problem). • In any case it is unlikely that this is what our visual system does, for many reasons – e.g., we do not in general find it more difficult to perceive a scene that has many identical parts, as would be predicted from this technique (since it would then be more difficult to find a unique descriptor for each object and the correspondence problem would quickly grow in complexity).

21. In virtue of what visual properties are objects individuated? • The most plausible property used in selecting and accessing an object is its location (this is often the only unique property available). The notion of a pointer suggests the use of location-as-access. • Virtually all theories of visual attention and property detection assume that we access an object’s properties by first retrieving its location.

22. But…. • Although there is a great deal of evidence for the priority of encoding location, this does not show that properties must be accessed by their location. • In studies in which objects remain stationary, location is confounded with individuality since in these cases being at a particular location is coextensive with being a particular individual. • But there is also recent evidence that we can access an object’s properties solely by virtue of the object’s persistence qua individual.This is referred to as object-based attention.

23. Unconfounding location and individuality • There are at least two possible ways to unconfound location and individuality: • use moving objects • use “objects” whose identity and/or ‘motion’ is independent of their spatial location.

24. Distinguishing access-by-location and access-by-individual • Moving objects • Object-specific priming (Object Files) • Object-specific Inhibition of Return * • Simultanagnosia & Visual Neglect * • Multiple Object Tracking (MOT) 2. Spatially coincident objects • Single-object advantage * • tracking in “feature space” * Some of these may be omitted for lack of time

25. Moving object studies… Object-specific Priming (object-file theory)Kahneman, D., Treisman, A., & Gibbs, B. J. (1992). The reviewing of object files: Object-specific integration of information. Cognitive Psychology, 24(2), 175-219. Sequence of displays in a simple Object-Priming experiment

26. Object File example: Wrong letter in box

27. Object File example: Correct letter in box

28. Moving object studies… Inhibition of Return(Tipper et al. 1991) When the target-cue interval is between 300 ms and 900 ms it takes longer to detect a cued target than an uncued one – this is called Inhibition of Return.

29. Moving object studies… “Inhibition of Return” moves with the object that is inhibited If the cued object moves, the Inhibition of Return moves with it.

30. Multiple Object Tracking Experiments How do we do it? What properties of individual objects do we use?

31. People can track 5 or more objects under a wide variety of conditions Objects don’t even have to avoid collisions!

32. Objects can even disappear from view, as long as they do it in the right way There must be local evidence of an occluding surface.

33. A possible location-updating tracking algorithm • While the targets are visually distinct, scan attention to each target and encode its location on a list. When targets begin to move; • For n=1 to 4; Check the n’th position in the list and retrieve the location Loc(n) listed there. • Go to location Loc(n). Find the closest element to Loc(n). • Update the n’th position on the list with the actual location of the element found in #3. This becomes the new value of Loc(n). • Move attention to the location encoded in the next list position, Loc(n+1). • Repeat from #2 until elements stop moving. • Go to each Loc(n) in turn and report elements located there. We compared the above algorithm with human performance on the very same displays. We assume (1) focal attention is required to encode locations (i.e., encoding is not parallel), (2) focal attention is unitary and has to be scanned from location to location. But it assumes no encoding (or dwell) time at each element.

34. Predicted performance of the location updating algorithm as a function of attention scanning speed

35. What properties are used in(a) selecting objects, and (b) tracking objects? Notice that these are different operations and need not involve the same properties

36. Role of object properties What properties can be used to select (index) an object in MOT? • We have evidence that under certain conditions selecting objects can be done either automatically or voluntarily. • Automatic selection requires “popout” features (sudden appearance, motion, stereo depth, etc) • Voluntary selection can use any discriminable property, but the objects must be attended serially and the property must be available long enough for this to occur (Annan study)

37. Role of object properties (continued)What properties can be used to track indexed objects? • We have some evidence that observers do not encode or use intrinsic object properties (e.g., color, shape) during tracking: • When we stop and ask, observers cannot tell us what properties objects had and they do not notice when properties like color/shape change during occlusion; • There is some evidence that tracking occurs (at least for small numbers of objects) even if it is not task-relevant (e.g., object-based priming and IOR); • We have some evidence that when objects differ in non-identifying (asynchronously changing) properties, they are not tracked any better than if they do not differ in these properties.

38. Role of object properties (continued)What properties can be used to track indexed objects? • We have some evidence that observers do not use an encoding of the trajectory of objects in tracking – i.e., tracking is not predictive (Brian Keane). • Tested condition in which all objects disappear for t milliseconds (up to half a second) then reappeared: • Where the would have been at that time (worst) • Where they were when they disappeared (best) • Where they were t ms previously (almost as good as above) • All shifted left, right, up or down by the same distance • Targets are tracked most poorly when they reappeared where they would have been at that time, best when they reappeared where they disappeared, and in between for the other conditions.

39. Role of object properties (continued)Do observers use some version of object locations for tracking? • It has been suggested that perhaps instead of using the location-updating method to track, observers respond to the objects’ “spatiotemporal trajectory” property (e.g., to their “space-time worms”).

40. Spatiotemporal continuity as a property that is used in tracking • Could a mechanism respond to spatiotemporal continuity without responding to object identity? • The notion of spatiotemporal trajectory presupposes that it is the trajectory of a single individual object, and not a sequence of time-slices of different objects. Therefore it assumes that the individual object has been selected and tracked. Responding to a spatiotemporal trajectory may be the same as tracking an object’s identity.

41. Another way to unconfound individuality and location • Can we attend to objects that are not distinguished by their location? • Single-object advantage studies • Can we track (generalized) objects that do not move through real space, but move through some other property space?

42. Observers can track non-spatial ‘virtual objects’ that move through a ‘property space’: Tracking superimposed surfaces Two superimposed Gabor patches that vary in spatial frequency, color and angle Blaser, Pylyshyn & Holcombe (2000)

43. Changing feature dimensions

44. Surfaces move randomly in “feature-space”

45. Snapshots snapshots taken every 250 ms Such generalized ‘objects’ can be tracked individually, and they also show single-object superiority for change detection.

46. Some speculations about what vision needs and what the Early Vision module may provide (1) 1. We need a mechanism that puts us in causal contact with distal objects in a visual scene – a contact that does not depend on the object satisfying a certain (conceptual) description, but on a brute causal connection. • We need such a connection in order to connect vision and action. • We need such a connection in order to ground concepts to their instances.*

47. Speculations on what vision needs and what the visual module may provide (2) • We need a mechanism that keeps track of the identity of distal objects without using their encoded properties – this happens whenever the correspondence problem is solved. • Such a mechanism realizes a rudimentary identity-tracker, with its own internal ‘rules’. 3. This is not a general identity-maintenance process; it will not allow you to recognize the identity of a person in a picture and a person on the street. But it may provide a way to maintain same-objecthood within the modular early vision system. There is also this tantalizing fact … • There is evidence for such a mechanism in babies as young as 4 months (Leslie, Spelke)!

48. Other studies: Implications for visually-controlled action, infant cognition, and robotics • A short tour of research in which the notion of deictic (or indexical) reference has been appealed to.

49. Ballard, Hayhoe et al.’s proposal for a “deictic strategy” People appear to use their direction-of-gaze as a reference point in encoding patterns and would prefer to make more eye movements rather than devote extra effort to memorizing a simple pattern.

50. Ballard, D. H., Hayhoe, M. M., et al. (1997). Deictic codes for the embodiment of cognition. Behavioral and Brain Sciences, 20(4), 723-767.